<?xml version="1.0" encoding="utf-8"?>
  <?xml-stylesheet type="text/xsl" href="rfc2629.xslt" ?>
  <!-- generated by https://github.com/cabo/kramdown-rfc version 1.7.43 (Ruby 3.4.12) -->


<!DOCTYPE rfc  [
  <!ENTITY nbsp    "&#160;">
  <!ENTITY zwsp   "&#8203;">
  <!ENTITY nbhy   "&#8209;">
  <!ENTITY wj     "&#8288;">

]>

<?rfc comments="yes"?>

<rfc ipr="trust200902" docName="draft-ietf-detnet-glbf-04" category="std" consensus="true" submissionType="IETF" tocDepth="5" tocInclude="true" sortRefs="true" symRefs="true">
  <front>
    <title abbrev="detnet-glbf">Deterministic Networking (DetNet) Data Plane - guaranteed Latency Based Forwarding (gLBF) for bounded latency with low jitter and asynchronous forwarding in Deterministic Networks</title>

    <author initials="T." surname="Eckert" fullname="Toerless Eckert" role="editor">
      <organization>Futurewei Technologies USA</organization>
      <address>
        <postal>
          <street>2220 Central Expressway</street>
          <city>Santa Clara</city>
          <code>CA 95050</code>
          <country>USA</country>
        </postal>
        <email>tte@cs.fau.de</email>
      </address>
    </author>
    <author initials="A." surname="Clemm" fullname="Alexander Clemm">
      <organization>Sympotech</organization>
      <address>
        <postal>
          <code>CA</code>
          <country>USA</country>
        </postal>
        <email>ludwig@clemm.org</email>
      </address>
    </author>
    <author initials="S." surname="Bryant" fullname="Stewart Bryant">
      <organization>Independent</organization>
      <address>
        <postal>
          <country>UK</country>
        </postal>
        <email>sb@stewartbryant.com</email>
      </address>
    </author>
    <author initials="S." surname="Hommes" fullname="Stefan Hommes">
      <organization>Rohde &amp; Schwarz</organization>
      <address>
        <postal>
          <country>DE</country>
        </postal>
        <email>stefan.hommes@icloud.com</email>
      </address>
    </author>

    <date year="2026" month="November" day="14"/>

    
    <workgroup>DETNET</workgroup>
    

    <abstract>


<?line 219?>

<t>This memo proposes a mechanism called "guaranteed Latency Based Forwarding" (gLBF) as part of DetNet for hop-by-hop packet forwarding with per-hop deterministically bounded latency and minimal jitter.</t>

<t>gLBF is intended to be useful across a wide range of networks and
applications that need high-precision deterministic networking services,
from in-car networks and networks for industrial automation on factory
floors all the way to country-wide networks at 100 Gbps and beyond.</t>

<t>Unlike other mechanisms, gLBF does not require network-wide clock
synchronization, nor does it require per-flow state on network nodes. It
thereby avoids the drawbacks of other known methods while retaining their
advantages.</t>

<t>Specifically, gLBF uses the queuing model and calculus of Urgency Based
Scheduling (UBS, <xref target="UBS"/>), which is used by TSN Asynchronous Traffic Shaping
<xref target="TSN-ATS"/>. gLBF is intended as a plug-in replacement for, or a mechanism
in parallel to, TSN-ATS, because it allows reuse of the same
controller-plane design that selects paths for TSN-ATS, sizes TSN-ATS
queues, calculates latencies and admits flows onto the calculated paths with
the calculated latencies.</t>

<t>In addition to reducing jitter compared to TSN-ATS through additional
buffering (dampening) in the network, gLBF also eliminates the per-flow,
per-hop state maintenance required by TSN-ATS. This avoids having to signal
per-flow state from the controller-plane to every hop, and the associated
scaling problems. It also reduces the implementation cost of high-speed
networking hardware, because it avoids the additional high-speed read/write
memory accesses needed to retrieve, process and update per-flow state
variables for a large number of flows.</t>



    </abstract>



  </front>

  <middle>


<?line 250?>

<section anchor="overview-informative"><name>Overview (informative)</name>

<section anchor="terminology"><name>Terminology</name>

<dl>
  <dt>CaaS</dt>
  <dd>
    <t>Control as a Service. Cloud (compute) and network services that enable control (loops) for network-connected devices, for example cars.</t>
  </dd>
  <dt>CQF</dt>
  <dd>
    <t>Cyclic Queuing and Forwarding. A queuing mechanism defined in Annex T of IEEE 802.1Q.</t>
  </dd>
  <dt>DT</dt>
  <dd>
    <t>Dead Time. A term from CQF indicating the time during each cycle in which no frames can be sent because the receiving node could not receive them into the desired cycle buffer.</t>
  </dd>
  <dt>node</dt>
  <dd>
    <t>A system that, with respect to gLBF, does not act as a host (sender/receiver). This memo avoids the term router so as not to imply an IP/IPv6 router as opposed to an LSR (label switching router). Likewise, a node can also be an IEEE 802.1 bridge implementing gLBF.</t>
  </dd>
  <dt>TCQF</dt>
  <dd>
    <t>Tagged Cyclic Queuing and Forwarding. A mechanism proposed to DetNet, see <xref target="TCQF"/>.</t>
  </dd>
</dl>

</section>
<section anchor="summary"><name>Summary</name>

<t>This memo proposes a mechanism called "guaranteed Latency Based Forwarding" (gLBF) for
hop-by-hop packet forwarding with per-hop deterministically bounded latency and minimal jitter.</t>

<t>gLBF is intended to be useful across a wide range of networks and
applications that need high-precision deterministic networking services,
from in-car networks and networks for industrial automation on factory
floors all the way to country-wide networks at 100 Gbps and beyond.</t>

<t>At its foundation, gLBF addresses the problems of burst accumulation and
jitter accumulation across multiple hops to achieve per-hop fixed latency
guarantees. This allows easy calculation of end-to-end latency as the sum of
these per-hop latencies.</t>

<t>In particular, gLBF eliminates the need for the controller-plane to interact
with the intermediate-hop routers of a gLBF path. This simplifies the
controller-plane and network operations, and increases the number of flows
the network can support: with gLBF, per-flow state is only required on
ingress routers. Although the controller-plane in general is out of scope
for this document, <xref target="controller-plane"/> discusses aspects relevant to system
design with gLBF, including the opportunity for a decentralized
controller-plane/admission control (<xref target="decentralized"/>).</t>

<t>Burst accumulation is the phenomenon whereby bursts of packets from senders
in an admission-controlled network grow across intermediate nodes. It can
only be managed with exponential complexity in admission control processing
and a significant increase in worst-case end-to-end latency and/or lower
maximum utilization. Dynamic, large-scale or easy-to-manage admission
control solutions need forwarding mechanisms without this problem, so that
admission control for bandwidth and jitter/buffer requirements can be
linear: decomposed into purely per-hop calculations that are independent of
changes in prior-hop traffic characteristics. This is not possible without
forwarding-plane solutions that overcome burst accumulation.</t>

<t>Existing solutions address burst accumulation by maintaining inter-packet
gaps on a per-flow basis, as in TSN Asynchronous Traffic Shaping (TSN-ATS).
gLBF instead ensures that inter-packet gaps are maintained without the need
for per-flow state. The basic idea is to assign a specific queuing delay
budget to each node and traffic class. This budget is known in advance from
the admission control calculus. When the packet is transmitted, the queuing
delay it actually experienced in the node is subtracted from that budget,
and the result is carried in a new header field of the packet. On receipt,
the next node passes the packet through a delay stage, where the packet
waits for the time indicated in that field before the node proceeds with
regular processing. In this way, queuing delay variations are absorbed, and
deterministic delay is achieved without the possibility of burst
accumulation.</t>

<t>By addressing burst accumulation, gLBF also overcomes the problem of jitter
accumulation, the second core problem of mechanisms such as TSN-ATS:
depending on the amount of competing admitted traffic on a hop at any point
in time, packets of a flow may experience anything from zero to the maximum
delay across the hop. This is the per-hop jitter. It adds up across multiple
hops, so applications requiring (near-)synchronous packet delivery cannot
rely solely on such mechanisms. It likewise limits how highly networks with
a large number of bounded latency flows can be utilized.</t>

<t>The basic principle on which gLBF operates was proposed by researchers in
the early 1990s under the name "Dampers". Dampers were not pursued or
adopted at the time because network equipment lacked the necessary
capabilities. Only recent and upcoming improvements in advanced forwarding
planes make it possible to build them at scale and at acceptable cost.</t>

<t>Unlike other proposals, gLBF does not require network-wide clock
synchronization. Likewise, it does not need to maintain per-flow state on
network nodes, because the delay budget and the queuing delay variations to
be absorbed are carried in the packets themselves, making them
"self-contained". This eliminates the per-flow, per-hop state maintenance
required by TSN-ATS, with its scaling problems of signaling per-flow state
from the controller-plane to every hop, and with the cost, in high-speed
networking hardware, of the high-speed read/write memory accesses needed to
retrieve, process and update these per-flow state variables for a large
number of flows.</t>

<t>Unlike other damper proposals, gLBF also supports the queuing model and
calculus of Urgency Based Scheduling (UBS, <xref target="UBS"/>), which is used by
TSN-ATS. gLBF is intended as a plug-in replacement for, or a mechanism in
parallel to, TSN-ATS, because it allows reuse of the same controller-plane
design that selects paths for TSN-ATS, sizes TSN-ATS queues, calculates
latencies and admits flows onto the calculated paths with the calculated
latencies.</t>

<t>While gLBF as presented here is intended for use with IETF forwarding
protocols, to provide DetNet QoS with bounded latency and lower, bounded
jitter, it would be equally applicable to other forwarding planes, such as
IEEE 802.1 Ethernet switching, assuming appropriate packet headers are
defined to carry the hop-by-hop and end-to-end metadata required by the
mechanism.</t>

<t>As described in this document, gLBF is a so-called admission-controlled
forwarding mechanism. It expects traffic that is allowed to be forwarded by
gLBF to be policed at the ingress of the network and to be forwarded through
the network along a path on which no link or queue can become congested.
Determining these conditions is the responsibility of an admission control
system outside the scope of this document. Typical elements providing this
complementary functionality are Path Computation Elements (PCE) and stateless
traffic steering mechanisms such as SRv6 or SR-MPLS. See also
<xref target="controller-plane"/>, in particular <xref target="path-steering"/>.</t>

</section>
<section anchor="application-scenarios-and-use-cases"><name>Application scenarios and use cases</name>

<t>gLBF addresses the same use cases that are targeted by deterministic
networking and high-precision networking services in general. Common
requirements of those services include the need to provide service levels
within very tightly defined bounds, in particular very specific latencies
without the possibility of congestion-induced loss. The ability to provide
services with corresponding service level guarantees enables many
applications that would simply not be feasible without such guarantees. The
following describes some use cases and application scenarios.</t>

<t>The development towards autonomous vehicles leads to new requirements and a
high demand for computational resources that are often not available within
a car. One solution is to reduce in-car processing to a minimum and to
offload computationally expensive tasks to the cloud. Because of the safety
implications of these use cases, delays in the transport of messages from
the cloud to the car and vice versa are often not acceptable. This includes
application scenarios such as trajectory planning for driverless vehicles,
but also emergency notifications to surrounding cars, which require fast
delivery of messages to avoid delayed reactions in case of an accident.
While the use of TSN for in-vehicle networks has already been investigated
and is more mature, real-time communication with guaranteed latencies
between vehicles and the cloud is still an open challenge.</t>

<t>Another important use case for the automotive industry is to further
optimize the manufacturing process by taking more data sources into account.
Since most SCADA systems and PLCs cannot connect to large data lakes or
perform more advanced computational jobs, real-time communication between a
cloud instance and the shop floor has several benefits. First, it integrates
more data into the decision-making process, since control algorithms for
automating manufacturing sites that run in a cloud environment can take more
variables into account than single-plant control systems. Second, it reduces
the resources needed on a particular factory floor or plant by transferring
complex, recurring and resource-intensive computational jobs to a cloud
provider, where computational power is not the limiting factor. However, in
such cases it is important that the associated control can still occur in
real time and according to very precise timing constraints. Such a
development is already foreseen by trends such as Industry 4.0 and
Control-as-a-Service (CaaS).</t>

</section>
<section anchor="background"><name>Background</name>

<t>The following background introduces and explains gLBF step by step.
It uses IEEE 802.1 "Time Sensitive Networking" (TSN) mechanisms for
reference, because at the time of writing, TSN bounded latency mechanisms
are the most commonly understood and deployed mechanisms providing bounded
latency and jitter services.</t>

<t>All mechanisms compared here, as well as those used by TSN, are based on an
overall service design in which traffic flows have a well-defined rate and
burstiness, which are tracked by the controller-plane and are called the
"envelope" of the flow in this memo.</t>

<t>The traffic model used for gLBF is taken from UBS. It gives the flow a rate
r [bits/sec] and a burst size b [bits], and the traffic envelope condition
is that the total number of bits w(t) of the flow over a period t [sec]
must be equal to or smaller than (r * t + b). In one typical case, a flow
sends one packet of size b every interval p [sec]. This translates into a
rate r = b / p for the flow because, after the flow has sent the first
packet of b bits, it takes p seconds until (r * t) reaches b again: (r * t)
= ((b / p) * p).</t>

<t>The controller-plane uses this per-flow information to calculate for each flow a path with sufficient free bandwidth
and per-hop buffers so that the bounded end-to-end latency of packets can be guaranteed. It then
allocates bandwidth and buffer resources to the flow so that further flows will not impact it.</t>

<t>Within this framework, bounded latency mechanisms can in general be divided into "on-time" and "in-time"
mechanisms.</t>

<section anchor="in-time-mechanisms"><name>In-time mechanisms</name>

<t>"In-time" bounded latency mechanisms forward packets in an (almost)
work-conserving manner.</t>

<t>When there is no competing traffic in the network, packets of traffic flows
that comply with their admitted envelope are forwarded without any queuing
latency introduced by the mechanism. When the maximum amount of admitted
traffic is present, packets of admitted flows experience the maximum
guaranteed, so-called bounded, latency. As a result, in-time mechanisms
introduce the maximum amount of jitter, because the amount of competing
traffic can change quickly, and with it the latency of packets.</t>

<t>IEEE TSN Asynchronous Traffic Shaping is the prime example of an in-time
mechanism. IETF Integrated Services (<xref target="RFC2212"/>) is an older mechanism
based on the same principles. In-time mechanisms have the benefit of not
requiring clock synchronization between nodes for their queuing.</t>

</section>
<section anchor="on-time-mechanisms"><name>On-time mechanisms</name>

<t>On-time bounded latency mechanisms deliver packets (near-)synchronously with
zero or a small maximum jitter, significantly smaller than that of in-time
mechanisms.</t>

<t>IEEE TSN Cyclic Queuing and Forwarding (CQF) is an example of an on-time
mechanism, as is the Tagged Cyclic Queuing and Forwarding (TCQF) mechanism
proposed to DetNet. Unlike the in-time mechanisms mentioned above, these
mechanisms require clock synchronization between routers.</t>

</section>
<section anchor="control-loops-vs-in-time-and-on-time"><name>Control loops vs. in-time and on-time</name>

<t>One class of applications that requires or prefers on-time (low-jitter)
delivery of packets is control loops in vehicles or industrial environments,
and hence in low-speed, short-range networks.</t>

<t>Emerging or future use cases such as remote PLC or remote driving extend
these requirements into metropolitan-scale networks. In these environments,
on-time forwarding is also called synchronous forwarding for synchronous
control loops.</t>

<t>Typically, in synchronous control loops, central units such as Programmable
Logic Controllers (PLC) control a set of sensors and actuators, polling or
periodically receiving status information from sensors and sending action
instructions to actuators. When packet forwarding is on-time (synchronous),
the central unit knows exactly when sensors sent a packet and when packets
are received by actuators, and can act accordingly.</t>

<t>These solutions do not require sensors and actuators to have accurate,
synchronized clocks. Instead, the central unit can control the time at which
sensors and actuators perform their operations, within the accuracy of the
(zero/low) jitter of the network packet transmission.</t>

<t>When bounded latency forwarding is (only) in-time, edge nodes in the network
and/or sensors and actuators need to convert the packets arriving with high
jitter into an on-time arrival model, to continue supporting the model the
application requires.</t>

<t>This conversion is typically called a playout buffer mechanism. It requires
time synchronization between senders and receivers, and hence most commonly
requires the network to support clock synchronization for these edge devices
and/or senders and receivers.</t>

<t>As a result, existing bounded latency mechanisms that support synchronous,
on-time delivery of packets require clock synchronization across the
network: in on-time mechanisms like CQF for the forwarding mechanism itself,
and in in-time mechanisms to synchronize edge devices and hosts.</t>

</section>
<section anchor="challenges-with-network-wide-clock-synchronization"><name>Challenges with network wide clock synchronization</name>

<t>Clock synchronization is a well-understood technology. However, it is also a
significant operational cost factor, and potentially an equipment cost
factor, for example through the difference in hardware cost between local
oscillators of different accuracy in network devices.</t>

<t>Clock synchronization, e.g., with IEEE 1588 PTP, is only deployed where no
simpler solution exists that provides the same benefits without it. Today,
for example, this means that in mobile networks only the so-called fronthaul
deploys clock synchronization, but not the backhaul.</t>

<t>As a result, it could be challenging to introduce new applications such as
the remote PLC or remote driving applications mentioned above if they had to
rely on a bounded latency network service with today's strict clocking
requirements. But even in existing markets such as in-car or industrial
networks, removing or reducing the need for clock synchronization could be a
significant step towards lower cost and simpler solutions. This reduction in
clock synchronization requirements is what gLBF provides.</t>

</section>
</section>
</section>
<section anchor="glbf-introduction-informative"><name>gLBF introduction (informative)</name>

<section anchor="dampers"><name>Dampers</name>

<t>The principle of the mechanism presented here is the so-called "Damper"
mechanism, first described in the early 1990s but never standardized,
primarily because the required forwarding was considered too advanced for
the equipment of the time. These limitations are now starting to be
resolved, so it seems like a good time to reintroduce the mechanism.</t>

<t>The principle of damper-based forwarding is easily explained: when node A
sends a packet, it has measured the latency L, i.e., how long the packet was
processed in the node. The main component of L is the queuing latency of the
packet in A caused by competing traffic to the same outgoing interface.
Based on the admission control and queuing algorithms used in the node,
there can be a known upper bound M(ax) for this processing latency. When the
packet then arrives at the next-hop node B, B simply delays the packet
further by (M-L). As a result, the packet has been forwarded synchronously
from A to B with a constant latency of M(ax).</t>

<figure title="Forwarding without Damper" anchor="FIG1"><artwork><![CDATA[
+------------------------+      +------------------------+ 
| Node A                 |      | Node B                 |
|   +-+         +-+      |      |   +-+         +-+      |
|-x-|F|---------|Q|------|------|-x-|F|---------|Q|------|
|   +-+         +-+      | Link |   +-+         +-+      |
+------------------------+      +------------------------+ 
  |<--------- Hop A/B latency --->|
]]></artwork></figure>

<t><xref target="FIG1"/> shows a single hop from node A to node B without Dampers.</t>

<t>A receives a packet. The F(orwarding) module determines the outgoing
interface towards node B, where the packet is enqueued into Q, because
competing packets may already be waiting to be sent from Q to B. This
happens because packets may arrive simultaneously from multiple input
interfaces on A (not shown in the figure), and/or because the speed of the
receiving interface on A is higher than the speed of the link to B.</t>

<t>Forwarding F can be L2 switching and/or L3 routing; the choice does not
impact the considerations described here.</t>

<t>When the packet is finally sent from Q over the link to B, B repeats the
same steps towards the next node (not shown).</t>

<t>x is the reception time of the packet in A and B. The per-hop latency of the
packet is xB - xA. It is predominantly determined by the time the packet has
to wait in Q, plus any other relevant fixed or variable processing latency
from F, plus the latency of the packet across the link. The latter is
predominantly determined by the serialization latency of the packet plus the
propagation latency of the link, which depends on the speed of light in the
material used; in fiber, for example, light travels 31 percent slower than
in vacuum.</t>

<t>All factors impacting the hop A/B latency other than Q in A are naturally
bounded: a well-defined maximum can be calculated or conservatively
estimated. The transmission latency of the packet from A to B is composed of
the serialization latency of the packet, which can be calculated from the
packet length and the bit rate of the interface, and the propagation latency
through the link, which can be calculated from the speed of light in the
link material and the length of the link.</t>

<t>To guarantee bounded latency through Q, an admission control system is
required that tracks, accounts for and limits the total amount of traffic
through Q. Admission control mechanisms rely on knowing the maximum burst
that each traffic flow can cause, and add up those bursts to determine the
maximum amount of simultaneous packet data in Q, which in turn determines
the maximum latency of each individual packet through Q.</t>

<t>With such an admission control system, one can therefore calculate a maximum
hop A/B latency MAX for a packet. However, packets will naturally have a
variable hop A/B latency lower than MAX, due to differences in packet size
and in competing traffic. As a result, their relative arrival times xB in B
will differ from their relative arrival times xA in A. This leads to
so-called "burst accumulation", and hence to the problem that the admission
control system cannot calculate the maximum burst and latency through Q in B
as easily as it can for Q in A.</t>

<t>If, however, packets could be forwarded such that their hop A/B latencies
were all the same (synchronous, on-time), then the admission control system
could apply the same calculus to B as to A. This is what damper mechanisms
aim to achieve.</t>

<figure title="Forwarding with Damper" anchor="FIG2"><artwork><![CDATA[
+------------------------+      +------------------------+ 
| Node A                 |      | Node B                 |
|   +-+   +-+   +-+      |      |   +-+   +-+   +-+      |
|-x-|D|-y-|F|---|Q|------|------|-x-|D|-y-|F|---|Q|------|
|   +-+   +-+   +-+      | Link |   +-+   +-+   +-+      |
+------------------------+      +------------------------+
        |<- A/B in-time latency ->|
        |<--A/B on-time latency ------->|
]]></artwork></figure>

<t><xref target="FIG2"/> shows the Damper mechanism that is simplest to explain, but not to
implement. Node A measures the time xA and sends this value in a packet
header to B. B measures the time xB directly on reception of the packet. It
then delays the packet for a time (MAX-(xB-yA)). As a result, the latency
(yB-yA) will be exactly MAX, up to the accuracy of the damper.</t>

<t>The first challenge with this simple approach is the need to synchronize the
clocks of A and B, so that (xB-yA) can be calculated correctly.</t>

<figure title="Forwarding with Damper and measuring" anchor="FIG3"><artwork><![CDATA[
+------------------------+      +------------------------+ 
| Node A                 |      | Node B                 |
|   +-+   +-+   +-+      |      |   +-+   +-+   +-+      |
|-x-|D|-y-|F|---|Q|----z-|------|-x-|D|-y-|F|---|Q|----z-|
|   +-+   +-+   +-+      | Link |   +-+   +-+   +-+      |
+------------------------+      +------------------------+
        |<- A/B in-time latency ->|
        |<--A/B on-time latency ------->|
]]></artwork></figure>

<t><xref target="FIG3"/> shows how this can be resolved by also measuring the time at z. A
calculates d = (MAX1-(zA-yA)) and sends d in a header of the packet to B. B
then delays the packet in D by d relative to the locally measured time xB.
Because neither the calculation of d in A nor the delay by d in D depends on
clock synchronization between A and B, this approach eliminates the need for
clock synchronization.</t>

<t>As a consequence of this change, however, the latency incurred by
transmitting the packet over the link is no longer included in the
measurement.</t>

<t>In this approach, MAX1 is the bounded latency for all processing of a packet
except the link propagation latency P, which is the length of the medium
divided by the speed of light in the medium, and MAX = MAX1 + P.</t>

<t>The serialization latencies across the sending interface on A and the
receiving interface on B can, however, be included in MAX1. It is only
necessary for the measured timestamps zA and xB to logically refer to the
same point in time, assuming a link of minimal length, e.g., a back-to-back
connection. For example, the reference time could be the time when the first
bit of the packet can be observed on the shortest wire between A and B. A
can most likely measure zA only some fixed time offset o before that
reference time, and hence needs to correct zA += o before calculating d.
Likewise, B may only be able to measure xB after the whole packet has
arrived. This would be l * r later than the reference time, where r is the
serialization rate of the receiving interface on B and l is the length of
the packet. B would hence need to adjust d -= l * r to subtract this time
from the time the packet needs to be delayed in D.</t>

<figure title="gLBF refined model" anchor="FIG4"><artwork><![CDATA[
+------------------------+      +------------------------+ 
| Node A                 |      | Node B                 |
| +-+    +-+   +-+   +-+ |      | +-+    +-+   +-+   +-+ |
|-|F|--x-|D|-y-|Q|-z-|M|-|------|-|F|--x-|D|-y-|Q|-z-|M|-|
| +-+    +-+   +-+   +-+ |      | +-+    +-+   +-+   +-+ |
+------------------------+      +------------------------+
       |<- Dampened Q ->|              |<- Dampened Q ->|
             |<- A/B in-time latency ->|
             |<--A/B on-time latency ------->|
]]></artwork></figure>

<t><xref target="FIG4"/> shows a further refined model for simpler implementation. Larger
routers/switches are typically modular, and forwarding happens on a
different modular component (such as a linecard) than the queuing for the
outgoing interface. Requiring clock synchronization even within such a large
device is undesirable.</t>

<t>Therefore, <xref target="FIG4"/> shows damper operations occurring solely before
enqueuing packets into Q and after dequeuing them. The Damper module
measures the timestamp x, extracts d from the packet header, adjusts it
according to the considerations described above, delays the packet by that
value and then enqueues it. After the packet is dequeued, the Marking module
measures the time z, adjusts it according to the same considerations,
calculates the value of d and rewrites it in the packet header before
sending the packet.</t>

</section>
<section anchor="dampers-and-controller-plane"><name>Dampers and controller-plane</name>

<figure title="Path selection and gLBF example" anchor="FIG5"><artwork><![CDATA[
         Controller-plane:
        Path Control  and
         Admission Control
      ..     .      .      ..
     .       .       .       . 
   +---+   +---+   +---+   +---+  Src---| A |---| B |---| C |---| D |
   +---+   +---+   +---+   +---+
     |       |       |       |
   +---+   +---+   +---+   +---+
   | E |---| F |---| G |---| H |--- Dst
   +---+   +---+   +---+   +---+
]]></artwork></figure>

<t>While the damper mechanism described so far can be realized with different
queuing mechanisms, and hence different calculi, on every interface that
supports bounded latency, the role of the controller-plane involves other
components besides admission control, and those need to be tightly coupled
with admission control for efficient use of resources.</t>

<t><xref target="FIG5"/> illustrates the most common problem. For a new traffic flow from
Src to Dst, the controller-plane needs to provide a path through the network
and reserve the resources along it. The simplest form, for bandwidth
reservation only, is called Constrained Shortest Path First (CSPF). When
end-to-end latency guarantees also need to be provided in support of bounded
latency, not only does the path calculation become more complex, but many
per-hop bounded latency queuing mechanisms also support more than one
per-hop latency on each hop, and the controller-plane needs to determine
which one is best for each hop of a candidate path.</t>

<t>As a result, to perform these operations, the controller-plane needs to know
exactly which parameters the bounded latency queuing mechanism on each
hop/interface supports. Standardizing these parameters ultimately means
standardizing the bounded latency queuing mechanism, and not only the damper
part.</t>

</section>
</section>
<section anchor="glbf-spec"><name>gLBF specification (normative)</name>

<section anchor="damper-with-ubs-queuingcalculus"><name>Damper with UBS queuing/calculus</name>

<t>Guaranteed Latency Based Forwarding (gLBF) uses the queuing model of Urgency
Based Scheduling (UBS), which is also used in TSN Asynchronous Traffic
Shaping (TSN-ATS). This allows one and the same controller-plane, including
its optimization algorithms for bandwidth and latency control, to be reused
or co-developed for both TSN-ATS (or a DetNet L3 equivalent) and gLBF.
Hence, gLBF should provide the most easily operationalized on-time solution
for networks that want to evolve from an in-time model based on TSN-ATS, or
that even want to operate both options in parallel.</t>

<t>Effectively, gLBF replaces the per-flow interleaved regulators of UBS with
per-flow stateless damper operations. Whereas UBS only provides in-time
latency guarantees, gLBF as a result provides on-time latency service, but
with fundamentally the same calculus as UBS.</t>

<figure title="UBS strict priority queuing block" anchor="FIG6"><artwork><![CDATA[
   +-----------------------------------------------+
   | UBS strict priority Queuing block             |
   |                                               |
   |               +--------------+   +----------+ |
   | +-------+   /-| Prio 1 queue |---|          | |
   | | Packet|  /  +--------------+   | Strict   | |
-->| | Prio  |->         ...         -| Priority | |--->
   | | enque |  \  +--------------+   | Schedule | |
   | +-------+   \-| Prio 8 queue |---|          | |
   |               +--------------+   +----------+ |
   +-----------------------------------------------+
]]></artwork></figure>

<t>Strict priority scheduling always takes the next packet from the
highest-priority queue (1 is highest) that has a pending packet, and
forwards it. UBS defines two options for the traffic model of traffic flows.
In the more flexible one, each flow specifies a rate r and a maximum burst
size b (in bits).</t>

<t>As a result, the bounded latency of a packet in priority 1 is (roughly) the
latency required to serialize the sum of the bursts of all the flows
admitted into priority 1. The bounded latency of a packet in priority 2 is
the priority 1 latency plus the latency required to serialize the sum of the
bursts of all the flows admitted into priority 2, and so on.</t>

</section>
<section anchor="glbf-processing"><name>gLBF processing</name>

<t>The following text extends/refines the damper processing as necessary
for gLBF.</t>

<figure title="Refined gLBF processing diagram" anchor="FIG7"><artwork><![CDATA[
   +------------------------+       +------------------------+ 
   | Node A                 |       | Node B                 |
   | +-+    +-+   +-+   +-+ |       | +-+    +-+   +-+   +-+ |
 --|-|F|--x-|D|-y-|Q|-z-|M|-|-------|-|F|--x-|D|-y-|Q|-z-|M|-|--->
IIF| +-+    +-+   +-+   +-+ |OIF IIF| +-+    +-+   +-+   +-+ |OIF
   +------------------------+       +------------------------+
          |<- Dampened Q ->|               |<- Dampened Q ->|
                |<- A/B in-time latency -->|
                |<--A/B on-time latency -------->|
]]></artwork></figure>

<t>The Damper module (D) retrieves the delay from a packet header field.
Requirements for that header field are defined in <xref target="DM"/>.</t>

<t>Because serialization speeds differ between Incoming InterFaces (IIF), and
because D will likely need to compensate the measured time x based on the
packet length and serialization speed, an internal packet header should
carry this information. Note that this is solely an implementation
consideration and should not impact the configuration model of gLBF.</t>

<t>The Queuing module is logically as described in UBS, except that the
priority of a packet in gLBF is not selected based on per-flow state, but
from an appropriate packet header field, which is looked up in the Packet
Prio Enque stage of Q; the packet is then enqueued according to that field's
value. Possible options for indicating such a priority in a packet header
field are defined below in <xref target="DM"/>.</t>

<t>Like Q, M needs to look up the packet priority to determine MAX1 for the
packet, and hence calculate the value d that it rewrites in the packet
header.</t>

<t>The overall configuration data model to be defined for gLBF is hence the
configuration data model of UBS, which is a set of priority queues and their
maximum sizes, plus, for gLBF, a maximum latency value MAX1 for each of
these queues, calculated by the controller-plane and used by the M module.</t>

<t>Given that the node knows the serialization speed of the OIF, MAX1 for each
queue could be derived automatically from the maximum length in bytes of
each UBS priority queue. Additional configuration parameters to fine-tune
the latency, such as increasing it by some percentage as a safety margin,
might be beneficial though. Such implementation details are outside the
scope of this specification.</t>

</section>
<section anchor="error-handling"><name>Error handling</name>

<t>A well-specified standard for a damper mechanism such as the one described
in this document also needs to handle error cases. The primary error
condition is when the sending node A of a gLBF packet recognizes that the
packet was enqueued for longer than MAX1, so that the delay to be written
into the packet would have to be &lt; 0.</t>

<t>In this case, error signaling such as ICMP/ICMPv6 needs to be triggered
(rate-limited!), and the packet needs to be discarded, to avoid admission
control failures or congestion for other packets further down the path.</t>

<t>The data model below describes an option that, instead of discarding the
packet, signals the condition through an ECN-like mechanism combined with
lower-priority forwarding of the packet. This informs the receiver directly
about the problem and allows the application to deal with such conditions
better than through other error signaling.</t>

</section>
<section anchor="DM"><name>Data Model for gLBF packet metadata</name>

<t>gLBF is specifically designed to support per-hop, per-flow stateless
operation, because its processing requires no per-flow metadata, only
per-packet metadata.</t>

<section anchor="damper-24-bits-or-more-unit-1-nsec"><name>damper: 24 bits or more, unit 1 nsec</name>

<t>This value is potentially different for every packet in a flow, and it is
rewritten by every gLBF forwarder along the path.</t>

<t>gLBF requires a packet header field indicating the delay that the damper on
the next hop needs to apply to the packet. Dampening only needs to be as
accurate as synchronous delivery needs to be. The maximum delay value is the
maximum per-hop queuing latency. This memo recommends a field of 24 bits or
more with a resolution of 1 nsec, which allows a maximum delay of ca. 16.7
msec. The reasons for this resolution and size are explained in
<xref target="damper-sizing"/>.</t>

<t>Note that link propagation latency does not affect the delay value. Hence,
the size of the delay field does not constrain the length of links.</t>

</section>
<section anchor="end-to-end-priority-3-bits"><name>end-to-end-priority: 3 bits</name>

<t>This field needs to be the same for all packets of a flow, so that they are
forwarded with the same latency processing and hence do not incur different
latencies and possible reordering. This field is written when the packet
enters the gLBF path and is only read afterwards.</t>

<t>Eight priorities, and hence eight different latencies per hop, are
considered sufficient. If the priority is an end-to-end packet header field,
such as 8 different DSCPs in IP/IPv6, this results in 8 different latency
traffic classes.</t>

</section>
<section anchor="hop-by-hop-priority-3-bits-per-hop"><name>hop-by-hop-priority: 3 bits (per hop)</name>

<t>This is a more advanced alternative to the end-to-end-priority. It is
optional. If it is present for a particular hop, it supersedes the
end-to-end-priority.</t>

<t>Because this data is per-hop, it should not be encoded in an end-to-end part
of the packet header, but in a hop-by-hop part of the header.</t>

<t>Because the resources of each DetNet flow need to be controlled, rerouting
DetNet flows without involving the controller-plane is highly undesirable
and could easily result in congestion. A per-hop, per-flow stateless
steering mechanism such as SR-MPLS or SRv6 is therefore highly desirable to
provide a per-hop steering field. The priority could easily be made part of
such a per-hop steering field by allocating, dynamically or on demand, up to
8 different SIDs (Segment IDentifiers), such as up to 8 MPLS labels, or by
using 3 bits of the argument (ARG) field of SRv6 SIDs.</t>

<t>When such a per-hop priority is indicated, the controller-plane can support
many more than 8 end-to-end latencies simply by combining different per-hop
latencies. For example, one flow may use priorities 1-1-1-1 across four
hops, the next, slower flow 1-2-1-2, the next one 2-2-2-2, and so on.</t>

</section>
<section anchor="phop-prio-3-bits"><name>phop-prio: 3 bits</name>

<t>This field is rewritten on every hop when hop-by-hop priorities are used.</t>

<t>It is an optional field that may be beneficial in an end-to-end header if
hop-by-hop priorities are used for forwarding in gLBF and the
hop-by-hop-priority of the packet on the prior hop is no longer available
from the hop-by-hop packet header. This is typically the case in MPLS-based
forwarding such as SR-MPLS, because this information will have been removed.
Note that this header field is only required to support optional simplified
high-speed forwarding implementations.</t>

</section>
<section anchor="error-handling-data-items"><name>Error handling data items</name>

<t>Di: one bit</t>

<t>Downgrade intent. This bit is set when the packet enters the gLBF domain and
is never changed.</t>

<t>Ds: one bit</t>

<t>Downgrade status. This bit is set to 0 when the packet enters the gLBF
domain and may be changed to 1 by a gLBF forwarder, as described below.</t>

<t>When the Di bit is not set and a gLBF forwarder recognizes that the packet
cannot be forwarded within its guaranteed latency, the packet is discarded
and forwarding-plane-specific error signaling (such as ICMP/ICMPv6) is
triggered. Discarding is done to avoid causing latency errors further down
the path because of this packet.</t>

<t>When the Di bit is set, the packet is not discarded; instead, the Ds bit is
set to indicate that the packet must no longer be forwarded with gLBF
guaranteed latency, but only as best-effort or even lower-than-best-effort
traffic.</t>

<t>Di and Ds may be encoded in the two ECN bits for IP/IPv6 data planes. The
required behavior should be backward compatible with existing ECN
implementations, in case the packet unexpectedly passes a router that does
not process it as gLBF.</t>

</section>
<section anchor="damper-sizing"><name>Accuracy and sizing of the damper field considerations</name>

<t>This memo recommends that the damper field in packet headers be 24 bits or
larger and represent the dampening time with a resolution of 1 nsec. The
following text explains the reasoning for this recommendation.</t>

<t>The accuracy needed for the damper value in the packet header, as well as
for the internal calculations performed for gLBF, depends on a variety of
factors. The most important factor is the accuracy of the bounded latency
desired/required by the applications.</t>

<t>Even though Ethernet is defined as an asynchronous medium, its clock
accuracy is required to be +/- 100 ppm (parts per million). For a 1500-byte
packet on 100 Mbps Ethernet, the serialization latency difference between
the fastest and slowest clock is 24 nsec. Because every gLBF node calculates
the damper value only from its own local clock and only for its own hop,
these errors do not compound across hops; in the worst case they add up
linearly, for example to 240 nsec across 10 such hops. On such lower-speed
networks with many hops, it is therefore likely not possible to provide
end-to-end latency accuracy much better than 1 usec without requiring more
accurate transmission timing through mechanisms such as PTP, which gLBF
attempts to avoid or minimize. For higher-speed links, such short-term
latency variations become irrelevant.</t>

<t>In WAN deployments, propagation latency is on the order of milliseconds,
e.g., ca. 1 msec for 250 km of fiber. Serialization latency on 1 Gbps
Ethernet is ca. 1 usec for a 128-byte packet. It is likely that a latency
accuracy on the order of 1 usec is sufficient when the round-trip time is
potentially 1000 times larger.</t>

<t>Based on these two simple data points, this memo considers an end-to-end
latency accuracy of 1 usec to be the target. To avoid introducing additive
errors, the resolution of the damper value needs to be higher. This memo
therefore proposes a resolution of 1 nsec for damper values, which is also
the resolution used in PTP.</t>

<t>The maximum value, and hence the size, of the damper field depends on the
maximum buffering latency in the sending node plus smaller factors such as
serialization latency. This maximum depends primarily on the slowest links
to be supported. A 128-byte packet on a 100 Mbps link has a serialization
latency of ca. 10 usec. With a 16-bit damper value at nsec resolution, this
would allow buffering of only 6 such packets, which may be too low. Hence,
the damper field should be at least 24 bits.</t>

</section>
</section>
<section anchor="ingress-and-egress-processing"><name>Ingress and Egress processing</name>

<section anchor="network-ingress-edge-policing"><name>Network ingress edge policing</name>

<t>When packets enter a gLBF domain from a prior hop, such as a node from a
different domain or a sending host, and that prior hop sends gLBF packet
markings, then the packet arrival timing and the gLBF markings in the packet
header(s) MUST only be honored when that prior hop is trusted by the gLBF
domain.</t>

<t>If the prior hop cannot be trusted to mark and/or time packets correctly,
the first-hop gLBF node in the gLBF domain MUST implement a per-flow policer
to prevent packets from such a prior hop from compromising gLBF guarantees
in the gLBF domain.</t>

<t>The policer is logically placed after the gLBF damper module and before the
queuing module.</t>

<t>Instead of a policer, the ingress edge node of the gLBF domain MAY use a
per-flow shaper or interleaved regulator. When, for example, a prior hop
sends packets of a flow at too high a rate, a shaper delays them further and
avoids discarding packets of that flow until the flow's burst size is also
exceeded. A policer would immediately discard any packet that does not
conform to its flow's envelope.</t>

</section>
<section anchor="glbf-sender-types"><name>gLBF sender types</name>

<section anchor="simple-glbf-senders"><name>Simple gLBF senders</name>

<t>Senders are expected to send their traffic flows such that their relative
timing complies with the admitted traffic envelope. Senders that send only
one flow, such as simple sensors or actuators, will typically be able to do
this.</t>

<t>When senders can do this (send gLBF packets with the right timing), these
packets do not need gLBF packet header elements. Instead, the ingress
network node can calculate the damper time locally, as described next,
avoiding any need for gLBF processing in the sender.</t>

<t>In this case, the maximum gLBF damper time is the serialization time of the
largest gLBF packet received from the sender. Thus, when the sender sends
packets smaller than the maximum admitted packet size, the first-hop network
node needs to dampen packets by the difference in serialization time.</t>

</section>
<section anchor="normal-glbf-senders"><name>Normal gLBF senders</name>

<t>When senders cannot fully comply with sending packets according to their
admitted envelope, they need to implement gLBF and indicate the gLBF damper
value in the packet header. For example, when the sender cannot ensure that
the outgoing interface is free at the target sending time of a gLBF packet,
another packet still being serialized will impact the timing of the
following gLBF packet(s). Another example is a sender that has to send
packets from multiple flows that cannot be, or are not, generated in a
coordinated fashion; these packets can then compete for the sender's
outgoing interface, introducing delay.</t>

<t>Normal gLBF senders simply need to implement the queuing and marking stages
of gLBF, but not the dampening stage, because that only applies to
receivers/forwarders.</t>

</section>
<section anchor="non-glbf-senders"><name>Non-gLBF senders</name>

<t>When senders can neither act as simple gLBF senders nor implement the gLBF
queuing and marking stages, the first-hop node of the gLBF domain needs to
treat them as an untrusted prior hop, but always use a shaper or interleaved
regulator so as not to discard their packets.</t>

</section>
</section>
<section anchor="receivers-and-glbf"><name>Receivers and gLBF</name>

<section anchor="normal-glbf-receivers"><name>Normal gLBF receivers</name>

<t>Normal gLBF receivers process gLBF packet markings and therefore need to
implement the gLBF dampening stage.</t>

</section>
<section anchor="glbf-incapable-receivers"><name>gLBF-incapable receivers</name>

<t>When receivers cannot implement the gLBF dampening stage, the worst-case
burst that may arrive at the receiver has to be counted as added jitter of
the end-to-end service.</t>

<t>The network cannot eliminate such bursts without additional coordination and
gating of packets across all senders and their network ingress nodes. This
is not a gLBF-specific issue, but simply a limitation of the
(near-)synchronous service model offered by gLBF, and it equally exists in
any delay and latency model with this type of service:</t>

<t>Consider a set of N senders that conspire to generate a burst at the same
receiver. They know from the admission control model the latency each of
them has to that receiver, and can thus time the generation of packets such
that the last-hop router would have to send all N packets (one from each
sender) on the interface simultaneously, which is obviously not possible.</t>

<t>With gLBF-capable receivers, this possible burst is accounted for by the
controller-plane including the latency introduced by such a worst-case burst
in the end-to-end latency, and by indicating to the receiver, in the gLBF
header field, the actual time the packet needs to be dampened. When the
receiver does not support such dampening, however, that maximum last-hop
burst simply turns into possible jitter.</t>

</section>
</section>
<section anchor="further-considerations"><name>Further considerations</name>

<t>A host may be a different type of gLBF sender than gLBF receiver. A common
case in cloud applications is a container or VM in a data center acting as
sender/receiver. Such an application may be considered trusted by the gLBF
domain if it belongs to the network operator itself (such as part of a
higher-level service of the network operator), but not if it is a customer
application.</t>

<t>gLBF marking needs to happen after contention between packets from all
applications, so it would typically happen inside the operating system. The
operating system of such a host may not (yet) implement gLBF, so it may not
be possible to correctly generate the gLBF marking as a sender. Instead, the
application may generate the appropriate packet markings for steering and
gLBF, but may have to leave the damper marking inaccurate.</t>

<t>Conversely, dampening of received gLBF packets on a receiver can happen at
the application level, so the operating system does not need to support it.
Such dampening is the same functionality that is normally called "playout
buffering" in applications, except that the delay is much smaller, because
playout buffering needs to take the whole path delay into account, whereas
gLBF dampening covers only the prior hop.</t>

</section>
<section anchor="summary-1"><name>Summary</name>

<t>There is a non-trivial set of ingress/egress processing options that could
help to simplify and secure the handling of different types of senders and
receivers.</t>

<t>Later versions of this memo may define specific profiles of edge behavior to
limit the recommended set of implementation options for senders/receivers
and gLBF edge nodes.</t>

</section>
</section>
</section>
<section anchor="controller-plane"><name>Controller-plane considerations (informative)</name>

<section anchor="glbf-versus-ubs-tsn-ats"><name>gLBF versus UBS / TSN-ATS</name>

<t>By relying on <xref target="UBS"/> for both the traffic model and the bounded latency
calculus, gLBF should be easy to operationalize using existing
controller-plane implementations for TSN-ATS: UBS/TSN-ATS and gLBF provide
the same bounded latency and use the same model to manage bandwidth for
different flows, namely calculating which flow goes into which priority
queue on which hop.</t>

<t>The main change to a UBS/TSN-ATS controller-plane when using gLBF concerns
what happens when a flow is admitted into the network, removed, or rerouted.
In UBS, each of these operations requires the controller-plane to signal the
traffic flow parameters to each node along the path, so that the forwarding
plane can establish the per-hop, per-flow state for the flow. Depending on
the network configuration, this also implies configuring the next hop of the
flow for routing.</t>

<t>For gLBF hop-by-hop operations, the first hop needs to receive the per-path
or per-hop priority information, which it then imposes in an appropriate
packet header of the flow's packets.</t>

</section>
<section anchor="first-hop-policing"><name>First-hop policing</name>

<t>In gLBF, the controller-plane has to perform this action only towards the
ingress node of the gLBF domain. Traffic parameters such as rate and burst
size are only needed there to establish a policer, so that the flow cannot
violate its admitted traffic parameters. This first-hop policing is actually
a function independent of gLBF, UBS or any other method used along the path.</t>

</section>
<section anchor="path-steering"><name>Path steering</name>

<t>gLBF operates independently of the path steering mechanism, but the
controller-plane will very likely want to ensure that gLBF (or, for that
matter, any DetNet) traffic does not unexpectedly get rerouted by in-network
routing mechanisms, because it normally will not, or cannot, reserve
resources for such rerouted flows on arbitrary failure paths without
significant additional effort and/or waste of resources.</t>

</section>
<section anchor="decentralized"><name>Decentralized controller-plane</name>

<t>Today's (pre-DetNet) controller-plane deployments commonly use one
centralized "SDN" controller to implement the functions of the "Path
Computation Element" (PCE), which tracks free bandwidth on every hop in the
network topology and can hence calculate on which path enough resources are
available for a requested new flow. This is, for example, the typical
deployment with RSVP-TE and Segment Routing.</t>

<t>However, it is equally possible to design a fully decentralized
controller-plane, especially for DetNet networks that may not have all the
complex policy requirements that make a centralized PCE most desirable. This
is possible because, with gLBF, the controller-plane does not need to
interact with non-ingress routers when dynamically allocating resources to
new traffic flows.</t>

<t>The model for a fully decentralized controller-plane could be derived from
the model developed and deployed for RSVP-TE without a centralized PCE,
where the ingress routers (LSRs) perform path calculations based on
IGP-flooded information about the available resources on every hop. However,
the necessary protocol operations may not lend themselves well to existing
IGPs, and a new decentralized controller-plane may even need to be
developed.</t>

<t>Note, though, that none of these (complex) controller-plane considerations
need to apply to deployments where the total amount of gLBF traffic is well
defined and so small compared to the available capacity that no bandwidth
exhaustion would occur, even under supported failure conditions. In those
cases, gLBF is still necessary to guarantee the latency of the traffic, but
the whole controller-plane is reduced to a network capacity planning and
ingress-router configuration task. This "controller-plane-free" mode of
operation is standard practice in many large service provider networks for
traffic that requires guaranteed bandwidth but has statically known
(limited) capacity requirements, such as IPTV.</t>

</section>
</section>
<section anchor="iana-considerations"><name>IANA considerations</name>

<t>None yet.</t>

</section>
<section anchor="acknowledgements"><name>Acknowledgements</name>

<t>Thanks to Florian Kauer for his input.</t>

</section>
<section anchor="open-issues"><name>Open Issues</name>

<t>[RFC-editor: please remove]</t>

<t><list style="numbers" type="1">
  <t>Data model: sketch a small YANG-like configuration data model (as in <xref target="TCQF"/>), primarily
including additional diagnostic elements such as the maximum occupancy of each priority
queue and the number of overruns.</t>
  <t>Collect details about the university Linux kernel implementation of gLBF mentioned in
the Status section of the appendix, and add a reference for it.</t>
</list></t>

</section>
<section anchor="changelog"><name>Changelog</name>

<t>[RFC-editor: please remove]</t>

<t>Individual document (draft-eckert-detnet-glbf)</t>

<t>00 Initial version.</t>

<t>01 Added use cases from Stefan</t>

<t>02 refresh only, waiting for progress in WG discussion</t>

<t>03 refresh, affiliation change of co-author</t>

<t>04 refresh, still waiting for more discussion in detnet</t>

<t>05 refresh, still waiting for more discussion in detnet</t>

<t>06 added Solution Scope, Taxonomy and Status for DetNet adoption considerations</t>

<t>07 Fixed author address (hommes)</t>

<t>WG document (draft-ietf-detnet-glbf)</t>

<t>00 Initial version</t>

<t>01 Just refreshing. Ran out of time to incorporate outstanding feedback for IETF126. Sorry.
   Will get to it within a month.</t>

<t>02 Updated author address (hommes). Changes in response to the WG adoption
   review by Florian Kauer:</t>

<t><list style="symbols">
  <t>Summary: added that with gLBF the controller-plane only needs to interact
with (and per-flow state is only needed on) ingress routers, with pointers
to the controller-plane considerations.</t>
  <t>New section "decentralized controller plane": centralized SDN/PCE
controller-plane as today's deployment model, option for a decentralized
controller-plane, and "controller-plane-free" deployments.</t>
  <t>Fixed "TSN-ATN" -&gt; "TSN-ATS" and "not starting" -&gt; "now starting".</t>
  <t>Resolved TBDs: clock synchronization cost factor, MAX1 fine-tuning.</t>
</list></t>

<t>Other changes:</t>

<t><list style="symbols">
  <t>Summary: per-hop fixed latency guarantees, end-to-end latency as sum of
per-hop latencies.</t>
  <t>Clock synchronization challenges: stated that gLBF reduces clock
synchronization requirements.</t>
  <t>Status: added the declarative traffic engineering implementation (<xref target="dgLBF"/>).</t>
  <t>Added Acknowledgements section.</t>
</list></t>

<t>03 With help of claude.ai: Editorial English fixes throughout
   (grammar, wording, typos, punctuation, US spelling, hyphenation). Semantic fixes:</t>

<t><list style="symbols">
  <t>Damper principle: delay is (M-L), not (B-L).</t>
  <t>Damper with UBS: gLBF provides on-time (not in-time) latency service.</t>
  <t>Link propagation latency P is medium length divided by speed of light (not times).</t>
  <t>High speed appendix: PIFO (not FIFO) hardware implementations are still research.</t>
  <t>PCE is "Path Computation Element".</t>
  <t>Terminology: TCQF is no longer called "the mechanism specified in this memo".</t>
  <t>Taxonomy: TCQF replaced by gLBF (copy-paste error).</t>
  <t>Fixed titles of the UBS strict priority queuing figure and the glbf_send() figure.</t>
  <t>Per-hop priority example: 1-1-1-1 across four hops (was five).</t>
  <t>IGMP -&gt; ICMP, PPT -&gt; PTP.</t>
  <t>Damper field accuracy: serialization (not propagation) latency in the 24 nsec and 10 usec examples.</t>
</list></t>

<t>04</t>

<t><list style="symbols">
  <t>Damper field accuracy: resolved the open question whether per-hop clock accuracy errors
add up across hops: they do not compound, but add up at most linearly (with example).</t>
  <t>Damper field: aligned the data model description (heading and text) with the 24 bits /
1 nsec recommendation of <xref target="damper-sizing"/>; it previously stated 1 usec units and 16 bits.</t>
  <t>High speed appendix: per-hop priority in the SRv6 SID ARG field is 3 bits (8 priorities),
consistent with the hop-by-hop-priority data model (was 4 bits, 16 priorities).</t>
  <t>Added Open Issues section: YANG-like data model (moved from inline TBD), details and reference for the university Linux kernel implementation (moved from inline TBD).</t>
  <t>References: CQF and TSN-ATS now point to the specific parts of IEEE Std 802.1Q-2022
(Annex T, and clauses 8.6.5.2.2, 8.6.5.6, 8.6.8.5 and 8.6.11) instead of the former amendments.
Removed the unused IEEE802.1Q and IEEE802.1Qbv references and a duplicate reference entry.</t>
  <t>Summary: added a paragraph stating up front that gLBF is an admission-controlled forwarding
mechanism relying on ingress policing, path computation (PCE) and stateless traffic steering,
with pointers to the controller-plane and path steering sections, in response to the WG
adoption review by Florian Kauer.</t>
  <t>Pseudocode fixes: glbf_enqueue() calculates the departure time from the arrival time plus
the damper value and stores it in the packet context for glbf_send(); glbf_dequeue() fixed
braces, dequeues from the selected queue and continues with the next packet; figure anchor
of glbf_send() now named after the function.</t>
</list></t>

</section>


  </middle>

  <back>


<references title='References' anchor="sec-combined-references">

    <references title='Normative References' anchor="sec-normative-references">



<reference anchor="RFC2210">
  <front>
    <title>The Use of RSVP with IETF Integrated Services</title>
    <author fullname="J. Wroclawski" initials="J." surname="Wroclawski"/>
    <date month="September" year="1997"/>
    <abstract>
      <t>This note describes the use of the RSVP resource reservation protocol with the Controlled-Load and Guaranteed QoS control services. [STANDARDS-TRACK]</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="2210"/>
  <seriesInfo name="DOI" value="10.17487/RFC2210"/>
</reference>
<reference anchor="RFC2212">
  <front>
    <title>Specification of Guaranteed Quality of Service</title>
    <author fullname="S. Shenker" initials="S." surname="Shenker"/>
    <author fullname="C. Partridge" initials="C." surname="Partridge"/>
    <author fullname="R. Guerin" initials="R." surname="Guerin"/>
    <date month="September" year="1997"/>
    <abstract>
      <t>This memo describes the network element behavior required to deliver a guaranteed service (guaranteed delay and bandwidth) in the Internet. [STANDARDS-TRACK]</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="2212"/>
  <seriesInfo name="DOI" value="10.17487/RFC2212"/>
</reference>
<reference anchor="RFC2474">
  <front>
    <title>Definition of the Differentiated Services Field (DS Field) in the IPv4 and IPv6 Headers</title>
    <author fullname="K. Nichols" initials="K." surname="Nichols"/>
    <author fullname="S. Blake" initials="S." surname="Blake"/>
    <author fullname="F. Baker" initials="F." surname="Baker"/>
    <author fullname="D. Black" initials="D." surname="Black"/>
    <date month="December" year="1998"/>
    <abstract>
      <t>This document defines the IP header field, called the DS (for differentiated services) field. [STANDARDS-TRACK]</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="2474"/>
  <seriesInfo name="DOI" value="10.17487/RFC2474"/>
</reference>
<reference anchor="RFC3270">
  <front>
    <title>Multi-Protocol Label Switching (MPLS) Support of Differentiated Services</title>
    <author fullname="F. Le Faucheur" initials="F." role="editor" surname="Le Faucheur"/>
    <author fullname="L. Wu" initials="L." surname="Wu"/>
    <author fullname="B. Davie" initials="B." surname="Davie"/>
    <author fullname="S. Davari" initials="S." surname="Davari"/>
    <author fullname="P. Vaananen" initials="P." surname="Vaananen"/>
    <author fullname="R. Krishnan" initials="R." surname="Krishnan"/>
    <author fullname="P. Cheval" initials="P." surname="Cheval"/>
    <author fullname="J. Heinanen" initials="J." surname="Heinanen"/>
    <date month="May" year="2002"/>
    <abstract>
      <t>This document defines a flexible solution for support of Differentiated Services (Diff-Serv) over Multi-Protocol Label Switching (MPLS) networks. [STANDARDS-TRACK]</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="3270"/>
  <seriesInfo name="DOI" value="10.17487/RFC3270"/>
</reference>
<reference anchor="RFC8655">
  <front>
    <title>Deterministic Networking Architecture</title>
    <author fullname="N. Finn" initials="N." surname="Finn"/>
    <author fullname="P. Thubert" initials="P." surname="Thubert"/>
    <author fullname="B. Varga" initials="B." surname="Varga"/>
    <author fullname="J. Farkas" initials="J." surname="Farkas"/>
    <date month="October" year="2019"/>
    <abstract>
      <t>This document provides the overall architecture for Deterministic Networking (DetNet), which provides a capability to carry specified unicast or multicast data flows for real-time applications with extremely low data loss rates and bounded latency within a network domain. Techniques used include 1) reserving data-plane resources for individual (or aggregated) DetNet flows in some or all of the intermediate nodes along the path of the flow, 2) providing explicit routes for DetNet flows that do not immediately change with the network topology, and 3) distributing data from DetNet flow packets over time and/or space to ensure delivery of each packet's data in spite of the loss of a path. DetNet operates at the IP layer and delivers service over lower-layer technologies such as MPLS and Time- Sensitive Networking (TSN) as defined by IEEE 802.1.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="8655"/>
  <seriesInfo name="DOI" value="10.17487/RFC8655"/>
</reference>
<reference anchor="RFC8200">
  <front>
    <title>Internet Protocol, Version 6 (IPv6) Specification</title>
    <author fullname="S. Deering" initials="S." surname="Deering"/>
    <author fullname="R. Hinden" initials="R." surname="Hinden"/>
    <date month="July" year="2017"/>
    <abstract>
      <t>This document specifies version 6 of the Internet Protocol (IPv6). It obsoletes RFC 2460.</t>
    </abstract>
  </front>
  <seriesInfo name="STD" value="86"/>
  <seriesInfo name="RFC" value="8200"/>
  <seriesInfo name="DOI" value="10.17487/RFC8200"/>
</reference>
<reference anchor="RFC8964">
  <front>
    <title>Deterministic Networking (DetNet) Data Plane: MPLS</title>
    <author fullname="B. Varga" initials="B." role="editor" surname="Varga"/>
    <author fullname="J. Farkas" initials="J." surname="Farkas"/>
    <author fullname="L. Berger" initials="L." surname="Berger"/>
    <author fullname="A. Malis" initials="A." surname="Malis"/>
    <author fullname="S. Bryant" initials="S." surname="Bryant"/>
    <author fullname="J. Korhonen" initials="J." surname="Korhonen"/>
    <date month="January" year="2021"/>
    <abstract>
      <t>This document specifies the Deterministic Networking (DetNet) data plane when operating over an MPLS Packet Switched Network. It leverages existing pseudowire (PW) encapsulations and MPLS Traffic Engineering (MPLS-TE) encapsulations and mechanisms. This document builds on the DetNet architecture and data plane framework.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="8964"/>
  <seriesInfo name="DOI" value="10.17487/RFC8964"/>
</reference>



    </references>

    <references title='Informative References' anchor="sec-informative-references">



<reference anchor="RFC3209">
  <front>
    <title>RSVP-TE: Extensions to RSVP for LSP Tunnels</title>
    <author fullname="D. Awduche" initials="D." surname="Awduche"/>
    <author fullname="L. Berger" initials="L." surname="Berger"/>
    <author fullname="D. Gan" initials="D." surname="Gan"/>
    <author fullname="T. Li" initials="T." surname="Li"/>
    <author fullname="V. Srinivasan" initials="V." surname="Srinivasan"/>
    <author fullname="G. Swallow" initials="G." surname="Swallow"/>
    <date month="December" year="2001"/>
    <abstract>
      <t>This document describes the use of RSVP (Resource Reservation Protocol), including all the necessary extensions, to establish label-switched paths (LSPs) in MPLS (Multi-Protocol Label Switching). Since the flow along an LSP is completely identified by the label applied at the ingress node of the path, these paths may be treated as tunnels. A key application of LSP tunnels is traffic engineering with MPLS as specified in RFC 2702. [STANDARDS-TRACK]</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="3209"/>
  <seriesInfo name="DOI" value="10.17487/RFC3209"/>
</reference>
<reference anchor="RFC4875">
  <front>
    <title>Extensions to Resource Reservation Protocol - Traffic Engineering (RSVP-TE) for Point-to-Multipoint TE Label Switched Paths (LSPs)</title>
    <author fullname="R. Aggarwal" initials="R." role="editor" surname="Aggarwal"/>
    <author fullname="D. Papadimitriou" initials="D." role="editor" surname="Papadimitriou"/>
    <author fullname="S. Yasukawa" initials="S." role="editor" surname="Yasukawa"/>
    <date month="May" year="2007"/>
    <abstract>
      <t>This document describes extensions to Resource Reservation Protocol - Traffic Engineering (RSVP-TE) for the set up of Traffic Engineered (TE) point-to-multipoint (P2MP) Label Switched Paths (LSPs) in Multi- Protocol Label Switching (MPLS) and Generalized MPLS (GMPLS) networks. The solution relies on RSVP-TE without requiring a multicast routing protocol in the Service Provider core. Protocol elements and procedures for this solution are described.</t>
      <t>There can be various applications for P2MP TE LSPs such as IP multicast. Specification of how such applications will use a P2MP TE LSP is outside the scope of this document. [STANDARDS-TRACK]</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="4875"/>
  <seriesInfo name="DOI" value="10.17487/RFC4875"/>
</reference>
<reference anchor="RFC8296">
  <front>
    <title>Encapsulation for Bit Index Explicit Replication (BIER) in MPLS and Non-MPLS Networks</title>
    <author fullname="IJ. Wijnands" initials="IJ." role="editor" surname="Wijnands"/>
    <author fullname="E. Rosen" initials="E." role="editor" surname="Rosen"/>
    <author fullname="A. Dolganow" initials="A." surname="Dolganow"/>
    <author fullname="J. Tantsura" initials="J." surname="Tantsura"/>
    <author fullname="S. Aldrin" initials="S." surname="Aldrin"/>
    <author fullname="I. Meilik" initials="I." surname="Meilik"/>
    <date month="January" year="2018"/>
    <abstract>
      <t>Bit Index Explicit Replication (BIER) is an architecture that provides optimal multicast forwarding through a "multicast domain", without requiring intermediate routers to maintain any per-flow state or to engage in an explicit tree-building protocol. When a multicast data packet enters the domain, the ingress router determines the set of egress routers to which the packet needs to be sent. The ingress router then encapsulates the packet in a BIER header. The BIER header contains a bit string in which each bit represents exactly one egress router in the domain; to forward the packet to a given set of egress routers, the bits corresponding to those routers are set in the BIER header. The details of the encapsulation depend on the type of network used to realize the multicast domain. This document specifies a BIER encapsulation that can be used in an MPLS network or, with slight differences, in a non-MPLS network.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="8296"/>
  <seriesInfo name="DOI" value="10.17487/RFC8296"/>
</reference>
<reference anchor="RFC8402">
  <front>
    <title>Segment Routing Architecture</title>
    <author fullname="C. Filsfils" initials="C." role="editor" surname="Filsfils"/>
    <author fullname="S. Previdi" initials="S." role="editor" surname="Previdi"/>
    <author fullname="L. Ginsberg" initials="L." surname="Ginsberg"/>
    <author fullname="B. Decraene" initials="B." surname="Decraene"/>
    <author fullname="S. Litkowski" initials="S." surname="Litkowski"/>
    <author fullname="R. Shakir" initials="R." surname="Shakir"/>
    <date month="July" year="2018"/>
    <abstract>
      <t>Segment Routing (SR) leverages the source routing paradigm. A node steers a packet through an ordered list of instructions, called "segments". A segment can represent any instruction, topological or service based. A segment can have a semantic local to an SR node or global within an SR domain. SR provides a mechanism that allows a flow to be restricted to a specific topological path, while maintaining per-flow state only at the ingress node(s) to the SR domain.</t>
      <t>SR can be directly applied to the MPLS architecture with no change to the forwarding plane. A segment is encoded as an MPLS label. An ordered list of segments is encoded as a stack of labels. The segment to process is on the top of the stack. Upon completion of a segment, the related label is popped from the stack.</t>
      <t>SR can be applied to the IPv6 architecture, with a new type of routing header. A segment is encoded as an IPv6 address. An ordered list of segments is encoded as an ordered list of IPv6 addresses in the routing header. The active segment is indicated by the Destination Address (DA) of the packet. The next active segment is indicated by a pointer in the new routing header.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="8402"/>
  <seriesInfo name="DOI" value="10.17487/RFC8402"/>
</reference>
<reference anchor="RFC8754">
  <front>
    <title>IPv6 Segment Routing Header (SRH)</title>
    <author fullname="C. Filsfils" initials="C." role="editor" surname="Filsfils"/>
    <author fullname="D. Dukes" initials="D." role="editor" surname="Dukes"/>
    <author fullname="S. Previdi" initials="S." surname="Previdi"/>
    <author fullname="J. Leddy" initials="J." surname="Leddy"/>
    <author fullname="S. Matsushima" initials="S." surname="Matsushima"/>
    <author fullname="D. Voyer" initials="D." surname="Voyer"/>
    <date month="March" year="2020"/>
    <abstract>
      <t>Segment Routing can be applied to the IPv6 data plane using a new type of Routing Extension Header called the Segment Routing Header (SRH). This document describes the SRH and how it is used by nodes that are Segment Routing (SR) capable.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="8754"/>
  <seriesInfo name="DOI" value="10.17487/RFC8754"/>
</reference>
<reference anchor="RFC8938">
  <front>
    <title>Deterministic Networking (DetNet) Data Plane Framework</title>
    <author fullname="B. Varga" initials="B." role="editor" surname="Varga"/>
    <author fullname="J. Farkas" initials="J." surname="Farkas"/>
    <author fullname="L. Berger" initials="L." surname="Berger"/>
    <author fullname="A. Malis" initials="A." surname="Malis"/>
    <author fullname="S. Bryant" initials="S." surname="Bryant"/>
    <date month="November" year="2020"/>
    <abstract>
      <t>This document provides an overall framework for the Deterministic Networking (DetNet) data plane. It covers concepts and considerations that are generally common to any DetNet data plane specification. It describes related Controller Plane considerations as well.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="8938"/>
  <seriesInfo name="DOI" value="10.17487/RFC8938"/>
</reference>
<reference anchor="RFC8986">
  <front>
    <title>Segment Routing over IPv6 (SRv6) Network Programming</title>
    <author fullname="C. Filsfils" initials="C." role="editor" surname="Filsfils"/>
    <author fullname="P. Camarillo" initials="P." role="editor" surname="Camarillo"/>
    <author fullname="J. Leddy" initials="J." surname="Leddy"/>
    <author fullname="D. Voyer" initials="D." surname="Voyer"/>
    <author fullname="S. Matsushima" initials="S." surname="Matsushima"/>
    <author fullname="Z. Li" initials="Z." surname="Li"/>
    <date month="February" year="2021"/>
    <abstract>
      <t>The Segment Routing over IPv6 (SRv6) Network Programming framework enables a network operator or an application to specify a packet processing program by encoding a sequence of instructions in the IPv6 packet header.</t>
      <t>Each instruction is implemented on one or several nodes in the network and identified by an SRv6 Segment Identifier in the packet.</t>
      <t>This document defines the SRv6 Network Programming concept and specifies the base set of SRv6 behaviors that enables the creation of interoperable overlays with underlay optimization.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="8986"/>
  <seriesInfo name="DOI" value="10.17487/RFC8986"/>
</reference>
<reference anchor="RFC9016">
  <front>
    <title>Flow and Service Information Model for Deterministic Networking (DetNet)</title>
    <author fullname="B. Varga" initials="B." surname="Varga"/>
    <author fullname="J. Farkas" initials="J." surname="Farkas"/>
    <author fullname="R. Cummings" initials="R." surname="Cummings"/>
    <author fullname="Y. Jiang" initials="Y." surname="Jiang"/>
    <author fullname="D. Fedyk" initials="D." surname="Fedyk"/>
    <date month="March" year="2021"/>
    <abstract>
      <t>This document describes the flow and service information model for Deterministic Networking (DetNet). These models are defined for IP and MPLS DetNet data planes.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="9016"/>
  <seriesInfo name="DOI" value="10.17487/RFC9016"/>
</reference>
<reference anchor="RFC9262">
  <front>
    <title>Tree Engineering for Bit Index Explicit Replication (BIER-TE)</title>
    <author fullname="T. Eckert" initials="T." role="editor" surname="Eckert"/>
    <author fullname="M. Menth" initials="M." surname="Menth"/>
    <author fullname="G. Cauchie" initials="G." surname="Cauchie"/>
    <date month="October" year="2022"/>
    <abstract>
      <t>This memo describes per-packet stateless strict and loose path steered replication and forwarding for "Bit Index Explicit Replication" (BIER) packets (RFC 8279); it is called "Tree Engineering for Bit Index Explicit Replication" (BIER-TE) and is intended to be used as the path steering mechanism for Traffic Engineering with BIER.</t>
      <t>BIER-TE introduces a new semantic for "bit positions" (BPs). These BPs indicate adjacencies of the network topology, as opposed to (non-TE) BIER in which BPs indicate "Bit-Forwarding Egress Routers" (BFERs). A BIER-TE "packets BitString" therefore indicates the edges of the (loop-free) tree across which the packets are forwarded by BIER-TE. BIER-TE can leverage BIER forwarding engines with little changes. Co-existence of BIER and BIER-TE forwarding in the same domain is possible -- for example, by using separate BIER "subdomains" (SDs). Except for the optional routed adjacencies, BIER-TE does not require a BIER routing underlay and can therefore operate without depending on a routing protocol such as the "Interior Gateway Protocol" (IGP).</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="9262"/>
  <seriesInfo name="DOI" value="10.17487/RFC9262"/>
</reference>
<reference anchor="RFC9320">
  <front>
    <title>Deterministic Networking (DetNet) Bounded Latency</title>
    <author fullname="N. Finn" initials="N." surname="Finn"/>
    <author fullname="J.-Y. Le Boudec" initials="J.-Y." surname="Le Boudec"/>
    <author fullname="E. Mohammadpour" initials="E." surname="Mohammadpour"/>
    <author fullname="J. Zhang" initials="J." surname="Zhang"/>
    <author fullname="B. Varga" initials="B." surname="Varga"/>
    <date month="November" year="2022"/>
    <abstract>
      <t>This document presents a timing model for sources, destinations, and Deterministic Networking (DetNet) transit nodes. Using the model, it provides a methodology to compute end-to-end latency and backlog bounds for various queuing methods. The methodology can be used by the management and control planes and by resource reservation algorithms to provide bounded latency and zero congestion loss for the DetNet service.</t>
    </abstract>
  </front>
  <seriesInfo name="RFC" value="9320"/>
  <seriesInfo name="DOI" value="10.17487/RFC9320"/>
</reference>

<reference anchor="I-D.ietf-detnet-dataplane-taxonomy">
   <front>
      <title>Dataplane Enhancement Taxonomy</title>
      <author fullname="Jinoo Joung" initials="J." surname="Joung">
         <organization>Sangmyung University</organization>
      </author>
      <author fullname="Xuesong Geng" initials="X." surname="Geng">
         <organization>Huawei</organization>
      </author>
      <author fullname="Shaofu Peng" initials="S." surname="Peng">
         <organization>ZTE Corporation</organization>
      </author>
      <author fullname="Toerless Eckert" initials="T. T." surname="Eckert">
         <organization>Futurewei Technologies</organization>
      </author>
      <date day="1" month="November" year="2026"/>
      <abstract>
	 <t>   This draft is to facilitate the understanding of the data plane
   enhancement solutions, which are suggested currently or can be
   suggested in the future, for deterministic networking.  This draft
   provides criteria for classifying data plane solutions.  Examples of
   each category are listed, along with reasons where necessary.
   Strengths and limitations of the categories are described.
   Suitability of the solutions for various services of deterministic
   networking are also mentioned.  Reference topologies for evaluation
   of the solutions are given as well.

	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-ietf-detnet-dataplane-taxonomy-07"/>
   
</reference>

<reference anchor="I-D.eckert-detnet-flow-interleaving">
   <front>
      <title>Deterministic Networking (DetNet) Data Plane - Flow interleaving for scaling detnet data planes with minimal end-to-end latency and large number of flows.</title>
      <author fullname="Toerless Eckert" initials="T. T." surname="Eckert">
         <organization>Futurewei Technologies USA</organization>
      </author>
      <date day="7" month="July" year="2025"/>
      <abstract>
	 <t>   This memo explain requirements, benefits and feasibility of a new
   DetNet service function tentatively called &quot;flow interleaving&quot;.  It
   proposes to introduce this service function into the DetNet
   architecture.

   Flow interleaving can be understood as a DetNet equivalent of the
   IEEE TSN timed gates.  Its primary role is intended to be at the
   ingress edge of DetNet domains supporting higher utilization and
   lower bounded latency for flow aggregation (interleaving of flows
   into a single flow), as well as higher utilization and lower bounded
   latency for interleaving occurring in transit hops of the DetNet
   domain in conjunction with in-time per-hop bounded latency forwarding
   mechanisms.

	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-eckert-detnet-flow-interleaving-03"/>
   
</reference>

<reference anchor="SCQF">
   <front>
      <title>Segment Routing (SR) Based Bounded Latency</title>
      <author fullname="Mach Chen" initials="M." surname="Chen">
         <organization>Huawei</organization>
      </author>
      <author fullname="Xuesong Geng" initials="X." surname="Geng">
         <organization>Huawei</organization>
      </author>
      <author fullname="Zhenqiang Li" initials="Z." surname="Li">
         <organization>China Mobile</organization>
      </author>
      <author fullname="Jinoo Joung" initials="J." surname="Joung">
         <organization>Sangmyung University</organization>
      </author>
      <author fullname="Jeong-dong Ryoo" initials="J." surname="Ryoo">
         <organization>ETRI</organization>
      </author>
      <date day="7" month="July" year="2023"/>
      <abstract>
	 <t>   One of the goals of DetNet is to provide bounded end-to-end latency
   for critical flows.  This document defines how to leverage Segment
   Routing (SR) to implement bounded latency.  Specifically, the SR
   Identifier (SID) is used to specify transmission time (cycles) of a
   packet.  When forwarding devices along the path follow the
   instructions carried in the packet, the bounded latency is achieved.
   This is called Cycle Specified Queuing and Forwarding (CSQF) in this
   document.

   Since SR is a source routing technology, no per-flow state is
   maintained at intermediate and egress nodes, SR-based CSQF naturally
   supports flow aggregation that is deemed to be a key capability to
   allow DetNet to scale to large networks.


	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-chen-detnet-sr-based-bounded-latency-03"/>
   
</reference>

<reference anchor="LSDN">
   <front>
      <title>Large-Scale Deterministic IP Network</title>
      <author fullname="Li Qiang" initials="L." surname="Qiang">
         <organization>Huawei</organization>
      </author>
      <author fullname="Xuesong Geng" initials="X." surname="Geng">
         <organization>Huawei</organization>
      </author>
      <author fullname="Bingyang Liu" initials="B." surname="Liu">
         <organization>Huawei</organization>
      </author>
      <author fullname="Toerless Eckert" initials="T. T." surname="Eckert">
         <organization>Huawei</organization>
      </author>
      <author fullname="Liang Geng" initials="L." surname="Geng">
         <organization>China Mobile</organization>
      </author>
      <author fullname="Guangpeng Li" initials="G." surname="Li">
         </author>
      <date day="2" month="September" year="2019"/>
      <abstract>
	 <t>   This document presents the overall framework and key method for
   Large-scale Deterministic Network (LDN).  LDN can provide bounded
   latency and delay variation (jitter) without requiring precise time
   synchronization among nodes, or per-flow state in transit nodes.

	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-qiang-detnet-large-scale-detnet-05"/>
   
</reference>

<reference anchor="TCQF">
   <front>
      <title>Deterministic Networking (DetNet) Data Plane - Tagged Cyclic Queuing and Forwarding (TCQF) for bounded latency with low jitter in large scale DetNets</title>
      <author fullname="Toerless Eckert" initials="T. T." surname="Eckert">
         <organization>Futurewei Technologies USA</organization>
      </author>
      <author fullname="Yizhou Li" initials="Y." surname="Li">
         <organization>Huawei Technologies</organization>
      </author>
      <author fullname="Stewart Bryant" initials="S." surname="Bryant">
         <organization>University of Surrey ICS</organization>
      </author>
      <author fullname="Andrew G. Malis" initials="A. G." surname="Malis">
         <organization>Malis Consulting</organization>
      </author>
      <author fullname="Jeong-dong Ryoo" initials="J." surname="Ryoo">
         <organization>ETRI</organization>
      </author>
      <author fullname="Peng Liu" initials="P." surname="Liu">
         <organization>China Mobile</organization>
      </author>
      <author fullname="Guangpeng Li" initials="G." surname="Li">
         <organization>Huawei Technologies</organization>
      </author>
      <author fullname="Shoushou Ren" initials="S." surname="Ren">
         <organization>Huawei Technologies</organization>
      </author>
      <date day="15" month="December" year="2025"/>
      <abstract>
	 <t>   This memo specifies a forwarding method for bounded latency and
   bounded jitter for Deterministic Networks and is a variant of the
   IEEE TSN Cyclic Queuing and Forwarding (CQF) method.  Tagged CQF
   (TCQF) supports more than 2 cycles and indicates the cycle number via
   an existing or new packet header field called the tag to replace the
   cycle mapping in CQF which is based purely on synchronized reception
   clock.

   This memo standardizes TCQF as a mechanism independent of the tagging
   method used.  It also specifies tagging via the (1) the existing MPLS
   packet Traffic Class (TC) field for MPLS packets, (2) the IP/IPv6
   DSCP field for IP/IPv6 packets, and (3) a new TCQF Option header for
   IPv6 packets.

   Target benefits of TCQF include low end-to-end jitter, ease of high-
   speed hardware implementation, optional ability to support large
   number of flow in large networks via DiffServ style aggregation by
   applying TCQF to the DetNet aggregate instead of each DetNet flow
   individually, and support of wide-area DetNet networks with arbitrary
   link latencies and latency variations as well as low accuracy clock
   synchronization.

	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-eckert-detnet-tcqf-12"/>
   
</reference>

<reference anchor="I-D.dang-queuing-with-multiple-cyclic-buffers">
   <front>
      <title>A Queuing Mechanism with Multiple Cyclic Buffers</title>
      <author fullname="Bingyang Liu" initials="B." surname="Liu">
         <organization>Huawei</organization>
      </author>
      <author fullname="Joanna Dang" initials="J." surname="Dang">
         <organization>Huawei</organization>
      </author>
      <date day="22" month="February" year="2021"/>
      <abstract>
	 <t>   This document presents a queuing mechanism with multiple cyclic
   buffers.

	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-dang-queuing-with-multiple-cyclic-buffers-00"/>
   
</reference>

<reference anchor="BIER-TE">
   <front>
      <title>Tree Engineering for Bit Index Explicit Replication (BIER-TE)</title>
      <author fullname="Toerless Eckert" initials="T. T." surname="Eckert">
         <organization>Futurewei Technologies Inc.</organization>
      </author>
      <author fullname="Michael Menth" initials="M." surname="Menth">
         <organization>University of Tuebingen</organization>
      </author>
      <author fullname="Gregory Cauchie" initials="G." surname="Cauchie">
         <organization>KOEVOO</organization>
      </author>
      <date day="25" month="April" year="2022"/>
      <abstract>
	 <t>This memo describes per-packet stateless strict and loose path steered replication and forwarding for &quot;Bit Index Explicit Replication&quot; (BIER) packets (RFC 8279); it is called &quot;Tree Engineering for Bit Index Explicit Replication&quot; (BIER-TE) and is intended to be used as the path steering mechanism for Traffic Engineering with BIER.

 BIER-TE introduces a new semantic for &quot;bit positions&quot; (BPs). These BPs indicate adjacencies of the network topology, as opposed to (non-TE) BIER in which BPs indicate &quot;Bit-Forwarding Egress Routers&quot; (BFERs). A BIER-TE &quot;packets BitString&quot; therefore indicates the edges of the (loop-free) tree across which the packets are forwarded by BIER-TE. BIER-TE can leverage BIER forwarding engines with little changes. Co-existence of BIER and BIER-TE forwarding in the same domain is possible -- for example, by using separate BIER &quot;subdomains&quot; (SDs). Except for the optional routed adjacencies, BIER-TE does not require a BIER routing underlay and can therefore operate without depending on a routing protocol such as the &quot;Interior Gateway Protocol&quot; (IGP).
	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-ietf-bier-te-arch-13"/>
   
</reference>

<reference anchor="DNBL">
   <front>
      <title>Deterministic Networking (DetNet) Bounded Latency</title>
      <author fullname="Norman Finn" initials="N." surname="Finn">
         <organization>Huawei Technologies Co. Ltd</organization>
      </author>
      <author fullname="Jean-Yves Le Boudec" initials="J." surname="Le Boudec">
         <organization>EPFL</organization>
      </author>
      <author fullname="Ehsan Mohammadpour" initials="E." surname="Mohammadpour">
         <organization>EPFL</organization>
      </author>
      <author fullname="Jiayi Zhang" initials="J." surname="Zhang">
         <organization>Huawei Technologies Co. Ltd</organization>
      </author>
      <author fullname="Balazs Varga" initials="B." surname="Varga">
         <organization>Ericsson</organization>
      </author>
      <date day="8" month="April" year="2022"/>
      <abstract>
	 <t>This document presents a timing model for sources, destinations, and Deterministic Networking (DetNet) transit nodes.  Using the model, it provides a methodology to compute end-to-end latency and backlog bounds for various queuing methods.  The methodology can be used by the management and control planes and by resource reservation algorithms to provide bounded latency and zero congestion loss for the DetNet service.
	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-ietf-detnet-bounded-latency-10"/>
   
</reference>

<reference anchor="I-D.stein-srtsn">
   <front>
      <title>Segment Routed Time Sensitive Networking</title>
      <author fullname="Yaakov (J) Stein" initials="Y. J." surname="Stein">
         <organization>RAD</organization>
      </author>
      <date day="29" month="August" year="2021"/>
      <abstract>
	 <t>   Routers perform two distinct user-plane functionalities, namely
   forwarding (where the packet should be sent) and scheduling (when the
   packet should be sent).  One forwarding paradigm is segment routing,
   in which forwarding instructions are encoded in the packet in a stack
   data structure, rather than programmed into the routers.  Time
   Sensitive Networking and Deterministic Networking provide several
   mechanisms for scheduling under the assumption that routers are time
   synchronized.  The most effective mechanisms for delay minimization
   involve per-flow resource allocation.

   SRTSN is a unified approach to forwarding and scheduling that uses a
   single stack data structure.  Each stack entry consists of a
   forwarding portion (e.g., IP addresses or suffixes) and a scheduling
   portion (deadline by which the packet must exit the router).  SRTSN
   thus fully implements network programming for time sensitive flows,
   by prescribing to each router both to-where and by-when each packet
   should be sent.

	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-stein-srtsn-01"/>
   
</reference>

<reference anchor="I-D.eckert-detnet-bounded-latency-problems">
   <front>
      <title>Problems with existing DetNet bounded latency queuing mechanisms</title>
      <author fullname="Toerless Eckert" initials="T. T." surname="Eckert">
         <organization>Futurewei Technologies USA</organization>
      </author>
      <author fullname="Stewart Bryant" initials="S." surname="Bryant">
         <organization>Stewart Bryant Ltd</organization>
      </author>
      <date day="12" month="July" year="2021"/>
      <abstract>
	 <t>   The purpose of this memo is to explain the challenges and limitations
   of existing (standardized) bounded latency queuing mechanisms for
   desirable (large scale) MPLS and/or IP based networks to allow them
   to support DetNet services.  These challenges relate to low-cost,
   high-speed hardware implementations, desirable network design
   approaches, system complexity, reliability, scalability, cost of
   signaling, performance and jitter experience for the DetNet
   applications.  Many of these problems are rooted in the use of per-
   hop, per-flow (DetNet) forwarding and queuing state, but highly
   accurate network wide time synchronization can be another challenge
   for some networks.

   This memo does not intend to propose a specific queuing solution, but
   in the same way in which it describes the challenges of mechanisms,
   it reviews how those problem are addressed by currently proposed new
   queuing mechanisms.

	 </t>
      </abstract>
   </front>
   <seriesInfo name="Internet-Draft" value="draft-eckert-detnet-bounded-latency-problems-00"/>
   
</reference>

<reference anchor="UBS" >
  <front>
    <title>Urgency-Based Scheduler for Time-Sensitive Switched Ethernet Networks</title>
    <author initials="J." surname="Specht" fullname="Johannes Specht">
      <organization></organization>
    </author>
    <author initials="S." surname="Samii" fullname="Soheil Samii">
      <organization></organization>
    </author>
    <date year="2016"/>
  </front>
  <seriesInfo name="IEEE" value="28th Euromicro Conference on Real-Time Systems (ECRTS)"/>
</reference>
<reference anchor="LBF" >
  <front>
    <title>High-Precision Latency Forwarding over Packet-Programmable Networks</title>
    <author initials="T." surname="Eckert" fullname="Toerless Eckert">
      <organization></organization>
    </author>
    <author initials="A." surname="Clemm" fullname="Alexander Clemm">
      <organization></organization>
    </author>
    <date year="2020" month="April"/>
  </front>
  <seriesInfo name="IEEE" value="2020 IEEE/IFIP Network Operations and Management Symposium (NOMS 2020)"/>
  <seriesInfo name="doi" value="10.1109/NOMS47738.2020.9110431"/>
</reference>
<reference anchor="gLBF" target="https://dl.ifip.org/db/conf/cnsm/cnsm2021/1570754857.pdf">
  <front>
    <title>gLBF: Per-Flow Stateless Packet Forwarding with Guaranteed Latency and Near-Synchronous Jitter,</title>
    <author initials="T." surname="Eckert" fullname="Toerless Eckert">
      <organization></organization>
    </author>
    <author initials="A." surname="Clemm" fullname="Alexander Clemm">
      <organization></organization>
    </author>
    <author initials="S." surname="Bryant" fullname="Stewart Bryant">
      <organization></organization>
    </author>
    <date year="2021" month="October"/>
  </front>
  <seriesInfo name="IEEE" value="2021 17th International Conference on Network and Service Management (CNSM), Izmir, Turkey"/>
  <seriesInfo name="doi" value="10.23919/CNSM52442.2021.9615538"/>
</reference>
<reference anchor="gLBF-Springer2023" >
  <front>
    <title>High Precision Latency Forwarding for Wide Area Networks Through Intelligent In-Packet Header Processing (gLBF)</title>
    <author initials="T." surname="Eckert" fullname="Toerless Eckert">
      <organization></organization>
    </author>
    <author initials="A." surname="Clemm" fullname="Alexander Clemm">
      <organization></organization>
    </author>
    <author initials="S." surname="Bryant" fullname="Stewart Bryant">
      <organization></organization>
    </author>
    <date year="2023" month="February"/>
  </front>
  <seriesInfo name="Springer" value="Journal of Network and Systems Management, 31, Article number: 34 (2023)"/>
  <seriesInfo name="doi" value="https://doi.org/10.1007/s10922-022-09718-9"/>
</reference>
<reference anchor="CQF" target="https://ieeexplore.ieee.org/document/10004498">
  <front>
    <title>IEEE Std 802.1Q-2022, Annex T: Cyclic queuing and forwarding</title>
    <author >
      <organization>IEEE 802.1 Working Group</organization>
    </author>
    <date year="2022" month="December" day="22"/>
  </front>
<annotation>Originally specified as IEEE Std 802.1Qch-2017, which has been incorporated into IEEE Std 802.1Q.</annotation></reference>
<reference anchor="TSN-ATS" target="https://ieeexplore.ieee.org/document/10004498">
  <front>
    <title>IEEE Std 802.1Q-2022, Clauses 8.6.5.2.2, 8.6.5.6, 8.6.8.5 and 8.6.11: Asynchronous Traffic Shaping</title>
    <author >
      <organization>IEEE 802.1 Working Group</organization>
    </author>
    <date year="2022" month="December" day="22"/>
  </front>
<annotation>Originally specified as IEEE Std 802.1Qcr-2020, which has been incorporated into IEEE Std 802.1Q.</annotation></reference>
<reference anchor="IPV6-PARMS" target="https://www.iana.org/assignments/ipv6-parameters/ipv6-parameters.xhtml">
  <front>
    <title>Internet Protocol Version 6 (IPv6) Parameters</title>
    <author >
      <organization></organization>
    </author>
    <date year="n.d."/>
  </front>
  <seriesInfo name="IANA" value=""/>
</reference>
<reference anchor="LDN" target="https://dl.ifip.org/db/conf/networking/networking2021/1570696888.pdf">
  <front>
    <title>Towards Large-Scale Deterministic IP Networks</title>
    <author initials="B." surname="Liu" fullname="Binyang Liu">
      <organization></organization>
    </author>
    <author initials="S." surname="Ren" fullname="Shoushou Ren">
      <organization></organization>
    </author>
    <author initials="C." surname="Wang" fullname="Chuang Wang">
      <organization></organization>
    </author>
    <author initials="V." surname="Angilella" fullname="Vincent Angilella">
      <organization></organization>
    </author>
    <author initials="P." surname="Medagliani" fullname="Paolo Medagliani">
      <organization></organization>
    </author>
    <author initials="S." surname="Martin" fullname="Sebastien Martin">
      <organization></organization>
    </author>
    <author initials="J." surname="Leguay" fullname="Jeremie Leguay">
      <organization></organization>
    </author>
    <date year="2021"/>
  </front>
  <seriesInfo name="IEEE" value="2021 IFIP Networking Conference (IFIP Networking)"/>
  <seriesInfo name="doi" value="10.23919/IFIPNetworking52078.2021.9472798"/>
</reference>
<reference anchor="multipleCQF" target="https://www.ieee802.org/1/files/public/docs2021/new-finn-multiple-CQF-0921-v02.pdf">
  <front>
    <title>Multiple Cyclic Queuing and Forwarding</title>
    <author initials="N." surname="Finn" fullname="Norm Finn">
      <organization></organization>
    </author>
    <date year="2021" month="October"/>
  </front>
</reference>
<reference anchor="dgLBF" target="https://doi.org/10.1016/j.future.2026.108494">
  <front>
    <title>Declarative traffic engineering for Low-Latency and reliable networking</title>
    <author initials="J." surname="Massa" fullname="Jacopo Massa">
      <organization></organization>
    </author>
    <author initials="S." surname="Forti" fullname="Stefano Forti">
      <organization></organization>
    </author>
    <author initials="F." surname="Paganelli" fullname="Federica Paganelli">
      <organization></organization>
    </author>
    <author initials="P." surname="Dazzi" fullname="Patrizio Dazzi">
      <organization></organization>
    </author>
    <author initials="A." surname="Brogi" fullname="Antonio Brogi">
      <organization></organization>
    </author>
    <author initials="A." surname="Clemm" fullname="Alexander Clemm">
      <organization></organization>
    </author>
    <author initials="T." surname="Eckert" fullname="Toerless Eckert">
      <organization></organization>
    </author>
    <date year="2026"/>
  </front>
  <seriesInfo name="DOI" value="10.1016/j.future.2026.108494"/>
<refcontent>Future Generation Computer Systems, Vol. 182, 108494, ISSN 0167-739X</refcontent></reference>


    </references>

</references>


<?line 1295?>

<section anchor="high-speed-implementation-considerations"><name>High speed implementation considerations</name>

<section anchor="high-speed-example-pseudocode-instead-of-reference-pseudocode"><name>High speed example pseudocode instead of reference pseudocode</name>

<t>This memo does not provide normative reference code aligned with the
normative functional description in <xref target="glbf-spec"/>. The primary reason is
that a naive 1:1 implementation of the functional description would lead to
inferior performance. Nevertheless, the feasibility of high-speed
implementations is an important factor in the deployability of mechanisms
like gLBF. Therefore, this appendix describes considerations for such
high-speed implementations, including pseudocode for one possible option.</t>

<t><xref target="gLBF"/> describes two possible optimized hardware implementation approaches
for gLBF: one using "Push In First Out" (PIFO) queues, the other using timed
FIFO queues. The following describes and explains the second approach,
because it is hopefully more feasible for nearer-term hardware
implementations, given that scalable high-speed PIFO hardware
implementations are still a matter of research.</t>

<t>However, the timed FIFO algorithm presented here may also face hardware
scalability challenges, because it requires O(IIF * priorities^2) FIFO
queues per outgoing interface, where IIF is the number of (incoming)
interfaces on the node and priorities is the number of desired priority
levels (TSN allows up to 8). If, however, the priority of flows (packets) is
not per-hop but the same on every hop (per-path), then the number is just
(IIF * priorities).</t>

</section>
<section anchor="ubs-high-speed-implementations"><name>UBS high speed implementations</name>

<t>The algorithm for the pseudocode shown later in this appendix is based on
the analysis of gLBF behavior performed in <xref target="gLBF"/>. That analysis reapplies
the same type of analysis and algorithm as <xref target="UBS"/>, which can be used to
implement TSN-ATS in high-speed switching hardware. Therefore, this appendix
starts by explaining the UBS mechanisms.</t>

<t>UBS (see <xref target="UBS"/>, Figure 4) logically consists of two separate stages. The
first stage is a per-priority, per-incoming-interface (IIF) interleaved
regulator. An interleaved regulator is a FIFO where dequeuing of the queue
head packet (qhead) is based on the per-flow state of qhead.</t>

<t>A target departure time for the qhead is calculated from the flow state that
the router maintains across packets, and once that departure time is
reached, the packet is passed to one of the egress strict priority queues.
It is called an interleaved regulator because the FIFO holds packets from
multiple flows (all from the same incoming interface and with the same
outgoing priority), hence 'interleaved', and 'regulator' because it delays
the packet until the target departure time.</t>

<t>The innovation of interleaved regulators, compared to the earlier per-flow
shapers as used in <xref target="RFC2210"/>, is the recognition and mathematical proof
that the arrival and target departure times of all packets received from the
same IIF (prior-hop node) with the same egress priority are in the same
order, so that these packets can be enqueued into a single FIFO in which
per-flow processing only needs to happen for the qhead, with no more than
negligible added latency compared to per-flow shapers.</t>

<t>With this understanding, the interleaved regulator and the following strict
priority stage can easily be combined into a single queuing stage with
appropriate scheduling. In this "flattened" approach, each logical egress
strict priority queue consists of the per-IIF interleaved regulators (FIFO
queues) for that priority.</t>

<t>Egress scheduling now needs to perform the per-flow state processing for the
qhead of each queue, and then take the target departure time of that packet
into account, selecting the interleaved regulator whose qhead has the
highest priority and earliest target departure time.</t>

</section>
<section anchor="glbf-compared-to-ubs"><name>gLBF compared to UBS</name>

<t>The same line of thought as in UBS can be applied to gLBF and is the basis
for the algorithm and pseudocode described here.</t>

<t>In gLBF, the order of packets by arrival and target departure time depends
on the egress priority and the IIF, but also on the priority the packet had
on the prior hop (pprio). All packets with the same (IIF,prio,pprio) arrive
in the same order in which they need to depart, so it is possible to enqueue
them into a FIFO and only perform timed dequeuing of the qhead.</t>

<t>Both prio and pprio matter because, when gLBF uses an encapsulation that
allows per-hop priority indications, a packet's priority can differ on each
hop. In UBS, the priority of a packet can equally differ across hops, but
the prior-hop priority is not needed to separate packets into interleaved
regulators, because UBS targets in-time delay, where the interleaved
regulator only needs to compensate for burst accumulation. In gLBF, in
contrast, the delay to be introduced depends on the damper value, which
depends on the prior-hop priority. As a result, the order in which gLBF
packets should depart (absent any burst accumulation) also depends on the
prior-hop priority.</t>

<t>Because gLBF does not deal with per-flow state on the router as UBS does in
its interleaved regulators, the FIFOs for gLBF are called timed FIFOs in
this memo rather than interleaved regulators.</t>

</section>
<section anchor="comparison-to-normative-behavior"><name>Comparison to normative behavior</name>

<t>Whereas the normative description defines D, Q and M blocks, where the D
block performs the dampening and the Q block performs the priority queuing,
the high-speed hardware optimization replaces D and Q with a single DQ
block. In it, packets are enqueued into a single-stage per-(IIF,pprio,prio)
FIFO (similar to UBS), and the dampening by the damper time happens on
dequeuing from that stage (also similar to UBS), while packets are
scheduled/prioritized based on their prio.</t>

<figure title="High Speed single stage gLBF" anchor="FIG-hs-diagram"><artwork><![CDATA[
   +------------------------+       +------------------------+ 
   | Node A                 |       | Node B                 |
   | +-+    +-------+   +-+ |       | +-+    +-------+   +-+ |
 --|-|F|--x-| DQ y  |-z-|M|-|-------|-|F|--x-| DQ y  |-z-|M|-|--->
IIF| +-+    +-------+   +-+ |OIF IIF| +-+    +-------+   +-+ |OIF
   +------------------------+       +------------------------+
          |<- Dampened Q ->|               |<- Dampened Q ->|
                 |<- A/B in-time latency ->|
                 |<--A/B on-time latency -------->|
]]></artwork></figure>

<t>In the normative definition, the damper value to be put into the packet
depends on the time y at which the packet leaves the damper stage D and
enters the priority queue stage Q. Simplified, the damper value is (MAX1 -
(z - y)).</t>

<t>In the high-speed implementation, packets are not passed from a damper stage
queue to a priority queue; instead, they stay in the same queue. On the one
hand, this is one of the core reasons why this approach can support high
speed: it does not require two-stage enqueuing/dequeuing. On the other hand,
it also eliminates the ability to take a timestamp at time y.</t>

<t>To replicate the normative gLBF behavior with a single stage, the marking
stage M (where the new damper value is calculated) therefore needs to know y
by other means.</t>

<t>Fortunately, y is exactly the target departure time of the packet from the D
stage, and therefore also from the DQ stage, so it already needs to be
calculated on entry to DQ: y = (x + damper - &lt;details&gt;), as explained
below.</t>

<t>The pseudocode consists of glbf_enqueue() describing the enqueuing
into the DQ stage, and glbf_dequeue() describing the dequeuing
from the DQ stage. glbf_dequeue() (synchronously) calls 
glbf_send() which represents the marking stage M.</t>

</section>
<section anchor="pseudocode"><name>Pseudocode</name>

<section anchor="glbfenqueue"><name>glbf_enqueue()</name>

<figure title="Reference pseudocode for glbf_enqueue()" anchor="FIG-glbf-enqueue"><artwork><![CDATA[
void glbf_enqueue(pak,oif) {
  tdamp = pak.header.tdamp // damper value from packet
  prio  = pak.header.prio  // this hops packet prio
  pprio = pak.header.pprio // previous hops packet prior
  iif   = pak.context.iif  // incoming interface of packet

  ta = adj_rcv_time(now(),                  \[1]
                    pak.context.iif.speed,
                    pak.context.length)

  td = ta + tdamp // (dampened earliest) departure time
  pak.context.td = td
  pak.context.max1 = max1\[oif,prio]
  enqueue(pak, td, q(oif,iif,prio,pprio))
}
]]></artwork></figure>

<t>pak.header contains the parsed fields of the received gLBF packet; tdamp is
the damper value. pak.context is a node-local header of the packet. iif is
the interface on which the packet was received.</t>

<t>prio and pprio are the current and prior-hop priorities from the packet
header. If, for example, SRH (<xref target="RFC8754"/>) is used with gLBF, and SIDs are
formatted according to <xref target="RFC8986"/>, then the priority for each hop could be
expressed as a 3-bit SID ARG field (see <xref target="RFC8986"/>, Section 3.1),
indicating 8 different priorities per hop.</t>

<t>Parsing the prior-hop priority is not a normative requirement of gLBF, but a
specific requirement of the high-speed algorithm presented here, which
allows the use of single-stage timed FIFOs instead of PIFOs. This type of
parsing is not required by SRH, but it is a benefit of packet headers that,
like SRH, retain the prior hops' information, as opposed to the SR-MPLS
approach, where past-hop SIDs/labels are discarded.</t>

<t>now() takes a timestamp from a local (unsynchronized) clock at the time of
execution.</t>

<t>[1] pak.context.ta (time of arrival) is the calculated time at which the
first bit of the packet entered the incoming interface of the node.
adj_rcv_time() is defined in the next subsection. td is the target departure
time, calculated from the arrival time ta and the damper value from the packet.
It simply converts the relative damper value into an absolute timestamp.</t>

<t>max1[oif,prio] is the only gLBF-specific interface-level state maintained
(read-only) across packets. It is the maximum time calculated by admission
control in the controller-plane for each priority. It is copied into a
context header field here, even though it is only used later during
dequeuing, on the assumption that accessing state information is undesirable
in the time-critical dequeuing/serialization stage, whereas such state
lookups are very common for many forwarding plane features before enqueuing.</t>

<t>Finally, the packet is enqueued into a per-(oif,iif,prio,pprio) timed FIFO
queue.</t>

</section>
<section anchor="adjrcvtime"><name>adj_rcv_time()</name>

<t>As the reference time against which the damper value in the packet is
applied, this specification assumes the time when the first bit of the
packet is seen on the medium connected to the incoming interface.</t>

<figure title="Example pseudocode for adj_rcv_time()" anchor="FIG-adj-rcv-time"><artwork><![CDATA[
time_t adj_rcv_time(now, speed, length) {
  time_t now
  int speed, length
      
  return now - length * 8 / speed - FUDGE
}
]]></artwork></figure>

<t>In most simple node equipment, the primary variable latency between that
(first-bit) measurement point and the time at which the now parameter of
adj_rcv_time() can be taken is the deserialization time of the packet. It
can easily be calculated from the packet length, assumed to be known from
pak.context.length, and the bit rate of the incoming interface, known from
pak.context.iif.speed, plus a fixed, measured or calculated latency FUDGE
for other internal processing. In the example pseudocode, these are assumed
to be static.</t>

<t>If other internal factors in prior forwarding within the node create
significant variation, they may need to be compensated through additional
mechanisms. In this case, rather than as shown in glbf_enqueue(), the
reference time should be taken directly upon receipt of the packet, before
it enters the forwarding stage F, and then be used as the now parameter of
adj_rcv_time(). This is deliberately not shown in the code, to highlight the
simplest implementation option, which should be sufficient for most node
types.</t>

</section>
<section anchor="glbfdequeue"><name>glbf_dequeue()</name>

<figure title="Pseudocode for glbf_dequeue()" anchor="FIG-glbf-dequeue"><artwork><![CDATA[
void glbf_dequeue(oif) {
  while(1) {
    tnow = now()
    ftd = tnow + 1  // time of first packet to send
    fq = NULL      // queue of first packet to send
    foreach prio in maxpriority...minpriority {
      foreach iif in iifs {
        foreach pprio in priorities {
          // find qhead with earliest departure time
          q = q(oif,iif,prio,pprio)
          if (q.qhead) {
            td = q.qhead.td   // target (departure) time
            if (td < ftd) {
              ftd = td
              fq = q
            }
          }
        }
      }
      if (fq) { // found packet
        pak = dequeue(fq,oif)
        glbf_send(oif,tnow,pak)
        break // next packet, restart with highest priority
      }
    }
  }
}
]]></artwork></figure>

<t>glbf_dequeue() selects as the gLBF packet to send the enqueued packet with
the highest priority and the earliest departure time td (as calculated in
glbf_enqueue()), where td must also not be in the future, because otherwise
the packet still needs to be dampened.</t>

<t>The outer loop across this hop's prio finds the packet with exactly these
properties, dequeues it and sends it. Strict priority is expressed through
the term maxpriority...minpriority, which searches from maximum to minimum
priority.</t>

<t>The two inner loops simply look for the packet with the earliest
target departure time across all the (IIF,pprio) timed FIFO queues
for the same prio and OIF.</t>

<t>In ASIC/FPGA implementations, finding this packet is not a sequential
operation as shown in the pseudocode, but can be a single-clock-cycle
parallel compare operation across the td values of all queue qheads, at the
expense of chip area, similar to how TCAMs work.</t>

</section>
<section anchor="glbfsend"><name>glbf_send()</name>

<figure title="Reference pseudocode for glbf_send()" anchor="FIG-glbf-send"><artwork><![CDATA[
void glbf_send(oif,tnow,pak) {
  qtime = tnow - pak.context.td          // \[1]
  td = pak.context.max1 - qtime - FUDGE2 // \[2]
  pak.header.tdamp = td
  serialize(oif,pak)
}
]]></artwork></figure>

<t>glbf_send() represents the marking block M of the specification. It needs to
calculate the new damper value td and update it in the packet header before
sending the packet.</t>

<t>td needs to be the maximum time max1 that the packet could have stayed in
priority queuing, minus the time qtime that it actually stayed in the queue,
minus any additional time FUDGE2 that the packet still needs to be processed
by the router before its first bit appears on the sending interface. This is
calculated as shown in [1] and [2] of <xref target="FIG-glbf-send"/>.</t>

<t>The primary factor impacting FUDGE2 is whether this calculation and the
update of the td field in the packet header can be done directly before
serialization starts, or whether they may need to be done while another
packet is still being serialized on the interface. Taking the added latency
of a packet currently being serialized into account can, for example, be
implemented by remembering the time when that packet started to be
serialized and its length, and then using these to calculate FUDGE2.</t>

</section>
</section>
<section anchor="performance-considerations"><name>Performance considerations</name>

<t>PTP clock synchronization needs to capture the accurate arrival time of PTP
packets. It also measures the residence time of PTP packets between
receiving and sending them with nsec accuracy, and adds it to a PTP packet
header field (correctionField). Given that this needs to be
hardware-supported functionality, it is unclear whether there is a relevant
difference between supporting it for a few PTP signaling packets and for a
much larger number of gLBF data packets, or whether it is a good indication
of the feasibility of the packet processing steps that gLBF requires.</t>

</section>
</section>
<section anchor="history-and-comparison-with-other-bounded-latency-methods"><name>History and comparison with other bounded latency methods</name>

<t>Before this work, the best solutions for the requirements outlined in this
document were <xref target="TCQF"/> and <xref target="SCQF"/>, which have been proven feasible for
high-speed forwarding planes through vendor high-speed forwarding plane
implementations using low-cost FPGAs for cyclic queuing. On the other hand,
implementation analysis concluded that <xref target="UBS"/> was infeasible for the same
type of high-speed, low-cost hardware, due to the need for high-speed
flow-state operations, especially read/write cycles to update the state for
every packet at packet processing speeds.</t>

<t>While <xref target="TCQF"/> and <xref target="SCQF"/> are very good solutions, proven to work at high
speed and low cost, available for deployment and ready for standardization,
the need for network-wide, hop-by-hop clock synchronization (albeit at lower
accuracy than required by other mechanisms feasible at high speed and low
cost), as well as the need to find good network-wide compromise cycle times,
still makes planning and managing them in dynamic, large-scale and/or
low-cost networks more complex to operationalize than the authors would
like.</t>

<t>In parallel to the short-term operationalizable solutions <xref target="TCQF"/> and
<xref target="SCQF"/>, <xref target="LBF"/> was researched, exploring a wide range of options to
signal and control latency through advanced per-hop processing and new
latency-related in-packet header elements. Unfortunately, given the
flexibility of <xref target="LBF"/>, it was impossible to find a calculus for simple
admission control. Ultimately, <xref target="gLBF"/> was designed out of the desire for a
mechanism derived from <xref target="LBF"/> that also provides guaranteed bounded latency
and has the flexibility benefits of <xref target="UBS"/> with respect to fine-grained,
per-hop independent latency management, but hopefully still fits within the
limits of what can be implemented in high-speed, low-cost forwarding/queuing
hardware.</t>

<t>The feasibility of gLBF in high-speed forwarding hardware has yet to be
proven. The following considerations explain why the authors think it should
be quite feasible:</t>

<t>The total number of FIFOs a platform needs to support is comparable to the
number of queues already supported in the same type of devices today, except
that not all service provider core nodes may implement that many, because in
the absence of DetNet services there is no requirement for them.</t>

<t>Implementing timed FIFOs (if the proposed high-speed implementation approach
is used) is no more complex than implementing shapers, which have by now
also started to appear on high-speed (100 Gbps and faster) core routers,
primarily for high-speed virtual leased line services.</t>

<t>For the rewriting of packet header fields based on the calculated latency
the packet experienced in the node, IOAM-type functionality arguably already
provides evidence of feasibility in high-speed hardware forwarding planes.
Processing of PTP timing protocol packets has similar requirements.</t>

<t>Even if implementation challenges at high speed and low cost make gLBF a
longer-term option in those networks, low-speed (1/10 Gbps) implementations,
such as in manufacturing or cars, may be an interesting option to explore,
given that gLBF has the no-jitter benefits of <xref target="CQF"/> without the need for
cycle management or clock synchronization: the best of both worlds, <xref target="CQF"/>
and TSN-ATS.</t>

</section>
<section anchor="solution-scope-taxonomy-and-status-for-detnet-adoption-considerations"><name>Solution Scope, Taxonomy and Status for DetNet adoption considerations</name>

<section anchor="solution-scope"><name>Solution scope</name>

<t>gLBF is not proposed as a standalone solution, but in conjunction with
<xref target="I-D.eckert-detnet-flow-interleaving"/>, a per-flow mechanism used on the
ingress gLBF router to allow more flexible per-flow management of the
admission of flow bursts than presented in this document. The proposal is
split into two documents because <xref target="I-D.eckert-detnet-flow-interleaving"/> can
be combined with all "mostly synchronous" per-hop mechanisms, for example,
besides gLBF, also CQF and <xref target="TCQF"/>.</t>

</section>
<section anchor="taxonomy"><name>Taxonomy</name>

<t>Using the section numbers/titles of the taxonomy document
<xref target="I-D.ietf-detnet-dataplane-taxonomy"/>, gLBF can be classified as follows.
In summary, gLBF is very much like TCQF, except that it replaces cycles with
Dampers, thereby further reducing end-to-end jitter, reducing clock accuracy
requirements (only phase synchronization, not clock synchronization, is
required) and removing the complexity of end-to-end cycle management.
Compared to older Damper-based proposals, gLBF reuses the ATS calculus,
which allows reuse of ATS admission control and delay calculations, and a
simple implementation based on FIFOs with timed scheduling (like ATS).</t>

<t><list style="symbols">
  <t>4.1 Per Hop Dominant Factor for Latency Bound  <vspace blankLines='1'/>
The per-hop dominant latency bound is Category 3. The per-hop latency for
each class is determined by the configured burst size.</t>
  <t>5.1 Periodicity  <vspace blankLines='1'/>
gLBF is not periodic. Its fundamental mode of operation is not based on
cycles to manage bursts, but on per-hop, per-packet synchronous delivery,
which maintains burst sizes end-to-end across all hops.</t>
  <t>5.2.  Traffic Granularity  <vspace blankLines='1'/>
Formally, gLBF has class-level granularity. However, the statement in
<xref target="I-D.ietf-detnet-dataplane-taxonomy"/> (revision 04) that all packets
within a class are treated the same is questionable for gLBF (and for
Damper solutions in general): in Damper solutions like gLBF, packet
latency is calculated per packet (within the same class) based on the
previously experienced per-hop latency, in order to compensate for it.</t>
  <t>5.3.  Time Bounds  <vspace blankLines='1'/>
gLBF is per-hop bounded. It is Right Bounded by the configured burst size
for the hop and class of the packet. It is Left Bounded by the signaled
remaining latency needed to achieve synchronous forwarding. If the bound
were observed not on the wire, but at a specific forwarding step in the
router (after reception, before enqueuing), then the per-hop bound would
be extremely tight, limited only by implementation accuracy.</t>
  <t>5.4 Service Order  <vspace blankLines='1'/>
gLBF service order is solely arrival based.</t>
  <t><list style="numbers" type="1">
      <t>Suitable Categories for DetNet</t>
    </list>
gLBF is "6.5.  Class level non-periodic bounded category" according to the
categorization.</t>
</list></t>

</section>
<section anchor="status"><name>Status</name>

<t>gLBF primarily has an open-source, simulation-level validation
implementation, as described in <xref target="gLBF"/>. Additional implementations were
done in the Linux kernel by a university and by a second
university in conjunction with a Prolog-based admission control server
implementation (<xref target="dgLBF"/>).</t>

</section>
</section>


  </back>

<!-- ##markdown-source:
H4sIAAAAAAAAA+29e3Mb2ZEv+P/5FLXsiDXgBsCHHq2mx46hHmzTV1JLotq+
N8azjgJQJMsNVKGrCqTYan33zfxl5nlUAVLbdzbWO7F971gk63UeefKdv5xO
p25RL8vq+jTbdlfTJ851ZbcqTrPnRVc067Iq265cZK+L7q5ufqT7shFdoV/H
2fO8y7M3q7wqsml2vc2bvOqKYpm9zLuiWtxnT/OWfjuvm7u8WeLJ65dPz8fZ
Vd1k83pbLenqSu+9K7ubbFXfZX8vO/pullfLLG/vq8VNU1f1tuWH7DVltXtw
rcvn86a4Pc2WRVcV3fR6Nb9yy3pR5Wua0LLJr7ppWdAso+vTo4fujib//MX7
1y/euwWN57pu7k+ztlu6ctOcZl2zbbuTo6Nvj05c2zVFvj7NLl68P3f0Gw3z
b/mqruj190XrNuWpy7Ksqxfyu/y8LDbdzWn2iH9t64ZecdX66+39Ovl9Ua/X
RdXpH1y+7W7qht865av8n8zmfV00q6JtsxeLH4ums4t1Q3M533bbprgryux9
sbip6lV9XRZt9sPlmd3G8yi60+zk5OQoe0bfa/JV9uLDpqE33uX3dtui7Ggl
Lmlf8+zZijbYX6iXNIZnZ9m3j44eHYW/bulN9ET0pWKdlytaxK7490U7u8q3
s2Vh15qaCa1Yll3dDGd4tio+0PoSMTxbFet1MsPL+/Wm7mhy/QH9iqGstsu7
8vrfF/zSGb1t+OXLriBa67KnzT1NPfnwBQ1oU9D/hL/ra9v5v7fy3ByPzWgn
h6P5H/annV+9yqvsj0wAbfLRd/XNssj+z+xycUPv/7n/YTw3u8Fz/14uVvV2
ufPjz184V9XNOu/K24JJ6t35s5OT46PTzP98Yn9++M1D/fHByTdH+uOTx48e
2Y90IOzHbx/Tva6srnrvfnBy9K3++PDJN+HJbx/bjw+P7IN0/aF/34Mn/scn
du+3R8f+x5PH9ti39A3+8WL6fBaf7CWxpg1zpmmXfyAGsr63uwocF7vviljO
tCS2RWcpv2Uu6Oi+y2dvz09x9+KmqOzetpnOmaNNlXdNlXfRAy8vn7+WB34q
8+ranqATc11M20W+KvRPdO97//J0KN3ipyung1zyS37aFlsa0ZRZ43S9XXXl
ht6zuF+sysV0vr26KpqWZ/X04sW76fsXp2ER5mXRTLtimjc4IM9fP315Olii
4Sz4DiKmsqKpdm21e8V6j003TT2nk9Ri4X54egkGmKkM+YHmzzeJJCDyLZbb
FR1olgDvy3UxvSyqtmSSyS5pmnw9e9HdFA19KXB1MKyiIRbGNHaqdH3x4gXN
+eQJyY0X26Zel4umzp7VFS0LfbPI6ip7V+SrKX+H+AVNbN1moxfP3r2/HOMV
RCM0xBOiK/waOC3/588mSRvixX+aZZcbYjf+0OuZ/VN9k1cV8dbkau/hS3o4
X5dl79nL+qYoV3qJiejpebJ4fyyvb6ZvmmJRtiVNxqRqJE/rW1rLNzntT0c3
1tdNvl7ntBu/cumOiPfzj4cX5xdv7Jns+03R0CmuqxZC+FVe5dcFiyRhum25
XWej19+/usQLxvrKZV2eZsdHs+Pjo28P+erDb7558GTGt8y+pT8+fHCcLPrJ
EQveL677+1lPvn1W+vUePpulkmOPXKHL1/21zw7wp+wNHaRzVksuOxo4vijr
HW8DdJfvhvoPL9/rIm+ml5Ee8yfoN5ODL23NcXb8Db32gnlThf0gCZ2St+0Y
f+eyaG5L+nO0XaNnry9fjSfZxc/rsplk77fNj8V9b7tOHnx7/O0h3/jo5OHD
E96v49m3j48fPXrwJN2v4+mxyPmOmRqpDjddt2lPDw+Xq1l5VW5Ykh4u54cL
GuLhomrX+B9+8PD40TdHxN6fPPpmtlle/Wts+r6Dmor83fqAEsz0ctPQ9hcN
TfJBj3r47GafPbvMAv9Sklg/I43SH9jsPdHJ9lr2fbUqr3kjL6qpEt0fi5yn
QGd9QesQVOp91GQjPM0O/lRvGyah+iqlG2WMgW4m2YPjCY2KFGviJNV2Pefn
HzzMRjxP/ZRSkKeBusT+MwM4OvrmsCUucHIyPeL/+/ab4yfTb1NiekCX/lvQ
AcvyeOcP+PzSfcvsydHJ7PjtlCZ7QqtJMuJD9p4UVEjvTEU7diAYNgc7lkSU
Tn4pXpj9RY2w74hMNuminkyPT6YnJzsPaVkUxYfNqm6KGf8oh7VebHnDadeO
jh4+/FYOPIkzmsb3TXldErms7rOWJBsd8ILtsaw3vcUNTfD4m0l2d1MubrIb
umNeFBWt4qJuNjXJEXqM1Ku6/+DsgFfv/eXr6dn7y1+zgmR9bFuSs09mj2eP
ZsSoJvrjY/nhyewRVpN/Pj6mfY5tx/dk+V3Rsl/e5Jv/b65zw8tw9M+u88Wb
Pz+evjl796q/1BAuxFeIoZCdWq+yP5NOyQzrcTa6eHP7eEzSjpQKNrbbvRLr
7PXZ6c7FuLu7m5E6nGMVcuJX1xVM28Nyc/t4uvFv7v8++3DTrVdQiUitTkb8
vuaz0hJDZc36kjXrni8gaDI64iDBPityDyBzY02Itz8SuKPetZQTHnhhyreF
ux6dHH3zRMXqw29Ovvn2yUGyVgefE6OVf030oxepj799/OTJExapuwh62uNd
T8uKmNZ19rLc7rnj8obOCv0fqc3Vnlue3Wz5HX+h/9lzx5+JJFlmnVXX5Yok
WL7nvjd5vaqzV8Uyv14RjZhuPM36gyrI5upKovVXxHnLfQP7E23Suiyyl8X1
Nr9n0jGDacChX+kF48VvI158/lle3P/oa7J3s/OyqiI6+37R1SQwA73tPBTE
Gvh8QmQeXtFCtYeb7ZxGw8yixQ5Xxd30it4dLD+aCUlTUsRu6VFWpOj9y4Hm
evC8WLCnBkZVp4yvoN2oiqIxxeMlWb2xltoUtAlsOAQy27kA3hh6Rcc57+9B
vqg3dXJp6uUprWw3MIDguKiTa/rE+Ywo5Jrsd9KBek+dFyS9y0U+uEEffTPL
nuc//9x/7E3eNeXPZZ1cnHpd4SmZT/1HzoihVvREfG36z2gX/7g645mW2KZN
cUUMgfaLOYZ497LvikoNNeJS682WvaaqzU2yP9erWXb8hKTk8dGTh98+JCvg
8vJ1RrbuN9NvHnz7P/cx8+ffX4gZR3ce/n12hU8x+yKpihftNgFi9W/fg246
nWb5vCWaXJDu9P6mbLN1sa6zTUNkw7I9p98XZE+X7Tojzr4imXbwK9zKB+ZX
Jqm4YfWMVFxxUYPWb+rNdH4/pX/oKrToq57pRgYvLi9jQQKZ3PdS81nhG9ak
SIujeuYcfzyjybAbCXeTGJ4XGakrV9tVli+auuW53bGyT3O5LniAlSn89EqX
bzZ09tXo7m7yji7Te27YAbDxRkQyvOik8jay7ddO3FVTr2kc00XeJF8Iv/CK
lNVyS9tQ0iTofNdrISL6/1e0M3Vz765Wdd3Qg6sVjabI7vJ7npP6EaeYSHh7
R/RylH0338iX5sV9XS1pWX6oVuWPNFd254SNJdrEei1r2vCq7oi0f9qWjX+h
vH2xqhc/OlPhyp8xwgnd38iDZXiON4+9eFnLFjrPQt9Edy+LdpZddI6HUMxp
+27rctliTssmv5sTObS8GTLGH6v6rqKREsejm0jTIn7YFF1OK06LTLeUjcuX
t+wMv6YXO3cpGhtIRacFHZXfbxr+mgaxwsLQbYvtaosPql9MSNmpXwzm3A9P
LyfZx4/0z6dPY9P3iLi2TPM0BVKZP6vbuo8fVav+9GmWDUgzZ0rcrLbX07Ki
yW1W+UKcBUQXE1KC4zPo6BZWyuggEh3UE1PXJ7THC9bGeRvoan3X0qv4d5oZ
z70ltuaYXzU1PdtM4Ygl8mUFUMi7LVbFouPj2t0ITfp3t+XPtIT6q+NlJMLW
xaMNbvUscjwDMaLluqQXXWEUNWvAPAJ/+1I/wefc9a74F9FWXlT0pmWJg0Dv
aGg/FrwfGo1aEIPNGznaOjT6jljq9ly+cuKUxT4u8/WmYMIZc7iKv6xkqYSS
r9o6I6lL5xmz4juMlCfOOJLQ9DrH/uWshSrZGy3wSGYZeKkS9w3c2DxQXm4a
VO+AgEVgJfobRI8Ut0VzzxxzgsXl20ia14uSF8yxI5vfbS5fPlsyEayXzqJc
k7LCNCVsZVG34MhgZmTj0Hsi1nVDfJh4cZHSVDimYXGjF9Dn8uXhXVN2hWMJ
QkPOF+wQYZZC12Wf6OiSdLulV2/EXYIpbTcsWntcw93mDfQfocU8g+NefR88
eJDXTETYulwuV4VzX2Xf3zLjLe6yURT7GNOVr7L34NQceCNl9FmeX7pTtiV4
weUMqr+O1Yh6u8xGCxHh45hbe8Yuh4YogFU03bdsREx600o01ZgnXavoYNEC
LAsRCbhMSglvChF/w7MgRdKdfl79nWVngYV5obwsSCOFpWkeDV6cYDW/pZc/
f+84dpwv4eDn97DUErKjD7PwgagTlkpa65pYwxaHpsiJ13F8o+AvCOsj9fCK
7cKWBl+xWG2ZWxmx8BtIPhYlaJ4ZPkup1VIlC1/ATWsxjsH4iQ3x+ZHvyImd
cVxsWdDAz7IWKhRWfCLqQVOwad4xTfHRnQTZReJSdvOGiXzUMo9tDvW7zVgP
JpSciKaxHsQ7mLHQ2cnlXfR2PjmsY5ANe8jWt91Ed9QbVpFA13T95eU72v58
Tny5RdiEZy8300dfkti9K1si/FyXhB7BMaXl47cHL8e8KZfX0ZHl9/AcaUHe
C5W8z6+v6btfJJZAJKrOYayihhFTLwoSavxKkko4IJfb9TonVeP/CT2QSN79
/1rf/zta31mXQR7zEqrWJvJuuWyEQ0PSqQThNZlvm5aP0mK7ZsHMg+LVsTyQ
5O+yqmYRs6BqcSToBBCn91t6VX4I2+c86bQmKUVnKfL23usDWIsr4rHLaVdP
6Z+w+zLkdrumG1iHaMOHeioEWx8lv67RWfcEPPaX92Kf/EUkms0jUCikKf9l
XSxZAuOTcsyxcrl8hDUcnVnLJ5mdiPjeUAeLhUvt43wi7MtqQXLVdqgn+1yk
wYCftFviSA3ZgBip8MWeplGyQkbHyassdUUq5TWTgc2COMeK1G1Wo3YuCYmB
a5i5K7xtC02iXdDQnSwj/dXcq6w291/w6VO2LNvFFoSXg42zrroqWIuHjgRu
71Q3jeZCq7HaLk1K1Zjstiq7e9UQlsVCsmZIXV0OFvqQ1dK2Ff1HxfXHj8kz
pN4T0TwdEn+pR+SmqGqaV82SUMwXnBTsvLC0VqSqiJ2W1XVm9PblqR9U2PPr
hnZHT1FMWcFW4s112LY5K54cFFrKuhQfNnXFMoI2g7WVVfGBl6OMPuknu/EB
KgclHbooTCVadiM0CHniOB0xLvpt19Grloe02kRSRePW+YdyTYdw25UrtQhn
2fP7Kl+Xi0kWpVqwGcNnm98mMwgjtJ0iwbvaCue1Q2nCIZiqmDcTHehMedaE
ZTZzazecNhLbaNDEK2nBeObCxQ5FzbCTACe46jOONOoibzhljRcVghO6ymZL
dHrvOU3Ep1RYkNbMPN2ykZg58cCv2TwmVtSUtTxpPkC6yKyF7BOWJMYLS9E+
6MNtydqlTtmFBdGjGBYMX+fUAxpwsYN9E1m/+MAfYSHln1IBsIvdE2nDxFFL
G4Q5FRJ31/mG2QhbrcZd5nlbMs/CPL9kD2cjtZLotInIrujEk3ZaVC2tsM4m
/mSGT/Lq2qD0BAglCBcH+0n5HS9ogcEtMpKVOQ4yq3dgLbnFd0Lwb1ms8nuy
GUkDAy+C+guFDbaXbduK3qCbpffST+KrwNm7hV3InMCJxdSnSnM8zLK/EE8R
5iIz5RGSaGzXTKXLSey4cBgdrLFFt4VaRAyA/YX0taW3aXm0LHe2c7j1+CSJ
fUmLKqOdOLMkabVJcPPtpKg0pbyFVFQyoG4krk2Ci3R39SLIGGfZ95Vo8ht6
lSz/h04+vMmDQiET8ja5LC5vzHUxEQYa3efuclFSmmCDqF1ic6Pxy2jmBd1W
hNmCtRVLdSk0xTWL+4jhEROthGGQAjVJNzuDnanngV6az9u6mfPKs8aTanq6
/q0pNykNynklRkgc2HQo1zuET+/t0PEAhucu9kTYcU7UM36zMDCXPgd9iPgV
e7V4caL7I+7Zbome8zZEeYVXIXFJyCdfs3LJjzHvK8Ax4NDhbbADgMPPjIx5
XkUcsaZTycKOt23iRSH0IZzGdR7TKj/TwUACZf5cNHWmpqBKFCV1FYt8gb4W
2KN5ZngIqv3D77EkGthu+jopGx4tZESi3gvnBz9idj8dxyxLiZeGUcL/QpKB
eLID/ycGyv/UlSxnWF4MYqWmXsZ6Ji3CDU2f7Qd6wqvqoNOhS6Nv5IgHTY1s
kbEFK/SBqXEuyQJ6d23WOQhIFMmCCb4N1t+cFb+24BREVlfLCoeXfqexHX/7
7REtHgIlOFhk4GcHz9ll1rQHHL/BT9kdH1uIJyLeLWuQ7IGtN0weeRfOrrkD
TM3hxd7ArbnipV0q2+YTyjbnIt/kODusuBN/gY6KqKV4iIgYIYfWNJlbldaB
1cbKgoNspN3Pf4TnyotRNga35Wopzgf2eEIzgTKEU1hsOvXmtF3fWS5rSOfy
n/eVx14AGpd/A5QdGpyJtqH73CXu80niapGDolLI2PpeDtfVbh54HBhexPgD
N8YRW7fF6pa/R0upWvfaHdAfr6DHQgof6KHc5zbN9rpN3Q63qXp3+Nj0HZuw
MuA9xV9TX+E/4kD1ZhxvMxsVX3SDqvDb6ezM9jo73WedncFmjXZ6p8vTDV2e
CWXCqz0kUIgQtQj3hD/c3vBH9g+EP5x3ef9vRTaYG/3TkY3Bvrt/JrKRDSMb
7p+ObPSuuNgt8RfEsWSbmD0X7EBlNxPz1ngBeaw8VbyQ61sSPqcJQjTgDp66
W2Y9GmV9W1/KU7scZzDeJnZJ3TpgSnfw1BKLoLMJBVMlpvJPobjIMBNeOzG9
wkV+TJ8z7t2hbB+0W/Bxei3RawM7V0Wt6JtQwpw5tdnXRezp3jQAcx/yJCLb
dF10OVcXJJEYdrZ46mIvWMt+5kVTzo3ZJY4KI14yCuqpejh32e1ul1kKwc/6
DZOZKUlixahryzsj9XE5Ovio/H1T0zoHGWpOGSXyKkoP7b9HFezEHcT1T9d8
+Igeg2pQ1aSUVD/yyQOpq2oBm5FmSGZqx/qFT94Snt/iosR8vPbFDnj6NdJ3
8x1OB6eee1KRWyZOnFd2Fcm8og0gMXK/YV8vSRIV70LRMoiydeLfQBCLCOJq
Wy0kCMWfZzn2hqcqWRdiwr6wF43ePHshQZzWMsadbRINT6KDO5Tky3e3j3mt
Lt9NX715yRUDRQG+6nZ5tSBKgreROCYv/tQ+YE72s6CC0lKQKGzKWiUDLzQ7
+9SRnbpnweb8LcHdIHkfQvKJtRLLMn59z4e9w2sdufZmvJJr0jwS9wg2jTTJ
+Al2yRXBkxpxIr0pW5EMXLVwn9IHIIw7Gku3uvfBK3Citr+CuNWb6J6Bus8Y
XUrFfGTZt84HalWLrc5qj9wXxuj8RMAryXASul5GyyLjD7WUrQb9WMOs7ndE
DISHthI5YgWPT2uRJ74cobHUC87+U2YVorgJpyJirJN9hxTaRUJqFix5sLUo
2p0mZnJooarXbNbcFjecQt7SpPIlfCFs6iebjC84phZ62VqzobNFOFh0RmmV
6m2ziOmwvqLtkQjcbV6ucpssifWcmTgr9cFjpY4YiVFbnCQY7HDSSHxnu1am
5+orkrv5Mh2K+kCIE3FUMW9/bM2SlJq77KlqD15buCq6ewefvO2aXGqjZZ6I
5tyaUgyPDGtSYk6TycIuPe/fwaf8d3OplAXtEAm3eX99vKVhBq0cotbt3FjP
jmgQfy8QFoLYrSxxcNmwiYpsNdveiZtvNQ2AdlU1O/o23L1mCdCLm4YPHr+I
w9Cm25lJc5W3nfMWcDzzToOnskyiEC9UPlRYQpMIi0XJntCZ6jy8QLob7CWU
GNhUhx3sY06mzlesZd9bUvUtH+trqFKIjNDxq+EP5KS2CQ9gNYXhycW628qW
UcIHIVAZuMicPsav9kfCjCfZTfagkcXNijJb0xX7aondE3thXaISTYjIiIiC
HehGOt6BhXhejYxPjfLdK9FfbRt+2JHNTDbTz4X6PaotB/0k6m4GA6sxYnth
stBx7OTBI03ry+6ambvkHF+6q+2yy2dnzy1mLrN68/JZqx6MTNMReCDifsBL
V2Qst2zKkynBaRPyPW9dp6f/7/W83b/itq6504WsuCR7Ufj1bREO5MAm9rll
+4zeOifJc0Xa9Sw7LxtYZuIDvoYjw4UFiNIGRJZN1T7VVWOtnr9nvtZ8dV2T
qXazhu7vLNDKi5oselt2xtCarThylRaK6rYkOx5MlVWmjl0LPKAoSSXeDn5J
xaO4XhVQDzo/GN0W1iVYqZpI2hwydZzqVbq/akiKnz0IRY0M6wKyxxvvZ0ph
JnVVNDwZ1ZY+8DYttviTJhXL61FfK0xzuLfCfTF1p4KyMYdteveGDQmLV/Dw
4fMCY8IwZ9kf6Y5bGBeVAyMTMVZCMw7HB6uepjZFvvJKj2K9oLnwm5j2xM0E
aUWrLgo5DRy8StQceKLA3Igx0erQnHnhwU5dLCjLwG3Ysdwy+WJBybgI/PfC
jvHD2RHMZ00emubtNJ9asd+IM4vGous9JcvmGixWhHOQ73N/hQmnqSVVC2bN
B9pQOjNikBCxbHgo/O/MXXSSzBhZWQdSSutLdkOVwwHCLONYsWXybwotngiG
dey2I8bMfg1Ya8yh++ZjeJvL1QUPprOAukjSGB7EtqvrJaazJKu/ZgkRDSMo
9maBxsapJhmYZsbclrY+etwn/jFBIuZ0V6xWkhNQixQ3n9IEohcV4jhIlavB
bFZeuVM/gc9tMrtATPybnJY0x/unpqsyO8L2w3tPf2q94MSKNOLeFAt06JAC
wcLtBhOTrdQD4i9Mi8WBaSnwCZmNyh4m1e5sdOLCwURZ3pjtynypEp/6D08v
YZReE1G04Z05hu+a7K//MSdee9gWi/+UIWkwgh0i2Vwv/2dIOQxVCzLUYBK6
sg3Ht6s7OpmRR5vdJXejbpzMDDXSiB6WRCYdfY3H4dZ0vLzvAe4GIoM1r1Mj
HHXUZL+l27/O5mOEdGp28KnZyHxlouEG1+Lg8mV1LsB5iJmJNxDBxVvmYPpx
VcXAQyWnVRi6w3Y32e/p0UO628S7RD3lANFnr7oi+ruItUqW5IqlmQvjmGNR
wPg7SN2Nxm3Y+05sTmc5hk5Faik9kF8TRzi1C+732WiE0Yzp981YaWNAaZr3
zGFyczP6nEhJqfXuKclI5FinUgn8BtCd2i3vewmvXVMUIZYONcy8uwp9YHF4
TNxYx44kgihdQqMbQUUD2dLzZDcQt1xgL9IAvo/ceyOkDmtvI1AtSw/yHYsP
FlIkcjhDsGQH/1/EIsUSIaFRcoH3szwMNsp/oXEvSxaPmh9wQLoIc9EDDPOg
1N9cFCFiufAVVw+D20bc1B1c2MOfGYB6ffzqSX7JKF8xEx47SzgFdxMFp0Jm
nEWZxb9Y1VFwz052PyU6CuIlTNFheaFf3HtPZ9mEIKFnEczmgpvKLF+OF1pA
22boZaDnm5FrzUfILdskhCjtm96dU3p3ahqFtLEJNUSRyDjmGGhwEjkBdTcm
th2z7Iy9hBI5Z71msJV+NntGbf7WOIyzI+zqJ8VUJ2kktHDl4keubPCRjFIV
r8HJ4vQ31hO+mI1h8dSGp2FpyWK+6eRc7OhkL/SF6eUe5KDlZCrFq/n0aQyV
ihj0ahkXmjgvib1HywcwOXg6WEoRwGAmYh0gYRNxWIvdItSW9UJt3ghBwMyY
dtkY4ekx/H7HMbS/feYQqknsCWwYPtaT4RDZRpwDcszTglFAlILFceVY1klG
z9VwC5KN/WwOMOmib89tK9KNrftvldwdoYRfk19MyiVeHvZ2mGM8yzRQJQ7t
weYiu7lmnSqf1xwngxcmmqn3RHx+ky1xUTbVUvqRip/dEl3Zp3kGNnHH3ijk
8WBFBk48/XILCws6c2vPcpb/3VS2cJz4RzxXbr3xIqMoIzdDmvYbmZatpOXc
gC/RE/wVhBsnbDI33VRSls1Jwild7NtB7gbJOalADJ5CM1saUh9Jwr95+Yxv
09/YZ4T0/g8cadIM2sQNCIG2Ljre1VVJRpom8vnPS05NgSzBeA62SjE4XCu+
KOWoeyDk+JRGl1yyhKzkiKbH3I9WJ35Jcuck04zOjHNDwzrEYDzuZX1NpP3M
q0wcJHj5bBx8BqSTidZIlhUysGFndtucrFr6BK3JShbeiRaraeqh8IFDDds2
0bgsM9S/sNXUG/GdOfaUNNuF98/576kMHCbNlxFRRgsylqyseB2QncaSj17K
3InfZ0OBqprb6yFZwufEztP6CYjnaBmkhq6Soguzwlf3opG2cYLisk6yJXYu
K89ZLC6284nvTqIUCnYIMBcA4SFdUHKekklCUuoWepOWXfJsnrndHzWPlwiI
OAP7zvRDG5HIVzbZRszZD+mEjs1e7QXqLPlNUvkQFDM1bJDlk+zniK3osfGs
SVZwMYiIsVRDc5qJu3tWFoyh1SDu1CWpHZzwcetrLtjTb5n9YvN48SB30uLC
3pzo+0g32RaWWmDZ2GKR8tLETmzjojOtKZHRtBYAsANtjIGTBfJ7VhJVxU8D
rPY2h7Htkwaaeq3uLin7UVIV1pr4K5xn9PHudT51Yo/oUY2CmR/vj1Z2RVsy
HIKEoIPiWFg+7mc0DUlf0JFEBzxw2V3i5/NSM+TUWYTwlCmrHspnyG6uD/NG
747QNyfrFKsrkV1ltUvOI6ffn+NkxWRbaEO89DYvu4bkbEdCWlV/Qs492zlP
xPLhuYlcUp1hb97HrsnOS6jcxVnxnhkgvZ6oRryaQkyMdYnMe6kPC/ltfKez
O+NSP8uEhe+6vDIQE85A0lwj+YpRMlu/K1e3CzJe5VTTJtuDXeBJZaiv1mWd
7VkUorrZ9Wxi+SSkRB4/evIke/P+zcSXhniPnTh8q9ohlInCOA3egXSVOtVB
HAWpzZHvLb6So/v1Mr+fuGg5Jubdyk3nonms63kZB4IwIrzZ22MkRSu6fbty
MtJ2N1mQebXtvF+a/a38TP8QssywlBcL8KgjOVhxHCJNNETLdhGH/WfUq+Sp
nrqblZAY97T5mi4mOaX5gCH0Sk/V6OYF/Q0HqZpyoWyKjcZYiZtlT2kNiMjh
tfAMZ5034BOmGWn4NVFMjTcg0LOub1XL9EXYPuCP4PDOA+gXNj1UcGlbaBqZ
SEL1UIh6lGZJ9visnOrK7f5aqrwyVgBRlNRiKYUyh1FXqW4tHhwWCmuqq7jV
ogzbq9QzsSNxK6VUzZ49iK0seAP7iUhpDi7oljlTBixjZrik/UwcG+h5U6IM
KC611ZSnuJYyh6jllBspkK+TRFnJ+vX8SufVoTRY9DaEcaKk+EpSFFXcc/6R
Y+/b6lZcJXyM2qIwkZFn1+C1LAYQ5O85RKKkrMEKSzajgLr2NCNOoZBw/wqp
p6eipqIC4ExdvqbF4mSzG5bYC9eULBMnyUu6PCuID3JuNjKlgnrEi+c0kpgW
VUgSCWewwkeDwise8kvbeUuwjHwxLGWtsKPKzjLsG7TooQNOvZjgocQ2r2tf
eEPCpJi5p7HzZFhSwufHY9iFkOe2N4uJIG+Y3zXXshXSMQpFAM9ejfIP48wX
80VpGd4FZj4554s8OOLL2mLRWjyJi0LgHMYOPZ1kTy0pRpMrwqI789bSwoxe
TV+Oe062aHs8zlrwLCaOF8kFPuPVfCqcMpf4X47Uc78zmCRRIMBzvp7u+e9r
geD5zHU8/0v2GlSY9f/7xf7B9afD687u+tq+Ff/8S/hn93V5fPph+sv5L35Q
v7zVn/0/+65/6esvOVvwC1//3107es+/+b+RUrbJzg6f+p2iv/3hF/fxNPvq
/OK7YwHU+v1BD+iVlQzltZ8y5z5+5Hs/fWK3yR0yOhGJR8EKqENYBrKfZFvS
l0BVN8U98BQ5/+cj/+0xWz3bVeHz7lQLssPr/OH18s6OQr/6CfytQlKmxhLe
ev+wC6zCNHwupwn5MRmXTnnOLOY85vkWp0BkqLshZYRMRS88knfh5PLxpOOW
V4W4MPEOX9hdVpttF6aECsCzbMQaFq9zZUzmqrwmhjueWKVoLKwke175YvCV
hHXCO2kt2CgNntD0Qclhxcyciyjh3Hjay5MIgkFH8fIB3IT0l9+J2+CmRvhX
izCcxoU0ZAvJqfIvyGresyiWEm3dleFFRkuPIGcyXOaATbEpcknId+D0rA61
nj6MbwqhhMVlTvUh5N5yAhlieBqyj0fDK8iy4KnQa1oWPxBKbfbhaTbNPpzB
wpbwyZJrbcQl7SnbB2ZEqif82NHkmAT520S2G64o4CiP5Ej50m5BAKC9sKSZ
HZJFuPe5vqQX2Ii+GhWG8fLKVOlWODFYfn9+EgyyllvF8p5P2Ajg2M6vd97K
37bgv1TStT66YRS74lxXPRyO1Ex8GmL5d/zXKyKt1E6c6COkF3DibPbgmDcR
tVCt6Mt8LLja7jZfbLdrzZAQi7PVCKfp6Dc9fip7gnP1VmmF9TvOomMKdmp7
nPazHixyoUcsKnlAcigijlCh6R2cqccTXcq2xD6wPWsdy+xSMjwQShBwh1+z
XbYLw/FZbZCRPJt4GkiGZcgJWCizutI4hfKikPuwgwBcbMvHVLD/+3soAtzB
k4V9UgcZkRkry3UIkw+MRBvQ28nOTHzD0KHD4S0GidRzvgr7xyR1TUuPOFYv
5YshqyOEKFVj9YvwltS1wQeTMI6YtqxqeoehEpSUyQqYUt5LwBF/rqZYoPyG
66cySfNR5AWiGX++9Yj1462xWPOFnZJJyMulBU20HdumimS5i4cZER2GyaXJ
ZFVyokqvzvmtZhaoeb1/LyZIUFnkGphHSXPIyMh9WLp/hF+d/U+tDzO9xHuy
TKZLuoMdanWq+2TFAVMIXIVfPsmWWxhuwUcl8AUyT86jMWffwIgZ6u0lJIBA
n5o7mWUIBA+94qnDYOVb/rB85qkzsC11DFgau4us7mFV9UHs/lUzy+qjQ+bh
EI9CjowlzvqtGZCvnJf+QZTZ5d5wzZH2yPvNuyfcl2O5V7BD0w30vpPIymFy
stGWTW8TURfBOqVBBUG5iINCE3Pxjidiru02IxV9RQbA/qv78DpfLwhGneNf
2wlzuWg5YpwnWK4jQKB/OZMr/t9dJlf/ejC5nv8yvVfDaqfJtfP6l77eM7l2
fv2fXzsPc6t2FyjIXPbe5vrDL/37pnxf3b9P/gv22cke+2xgnJ1444xp63mP
ZDKrnBN3YCtIHOL2idy6tfMYaTMjBnX3tEFV/XDmQ62aDUe8ZFtIerd6HxTq
Qoylp7te8pQYFKndncixoID3cDEE17QaOjiUYUuklnjsdPTh6fT+bLzL0WEq
xugetwgznxc+dgsWzWKw3hWa1AOovjVxN/rSBcvX8msrhZi5FPR6n24vYMNS
VWKv/A21LjzmTqZT2aH6oJ6Kxzz7b3zmf/7Cmf/5v/eZf/D5My9AfThQnJPu
OcADzwHYBwuKVPoxxzJSDTgm559OQvo/08lxEQTtMvs9TtbxdPTzGY5WdO4V
1kbPeWp76Knfd3C5sSFqK4NCogcP4bnVfeRiFlYxc0899kWp5lY4Fso1BLAT
6Mk4tILicC9/fx4Zk3vCHRYj9KcRS+hP8x6Qu90vs6AYzLifttLKSOtzJe8w
0lBim7ysuLZESpk9aFGZutNTN4gkn7LLHdkGqHtbeuNYFhIMHdh9yZwmzPiO
jU/tSKKA6hO5FIA+oxy++MAcO4xjl03/JsI2GNpgjMrGuDSa72uOhF02nd4r
Sidr67+XoX+dvVG2vMugBcZAcGtYblDPN6YG4j7n2VM+RdFuzYtklXkY5uhB
9oMHYPEx/oSa244OcZv9LN8lKchVY5w6pclOVyI1vTMLOECZxwEK5f5ad37l
MTtlcS0enSNAy1ni/K/TGjVgppynEeMi8wUsmRafqabsmcOdKbeSeT8vu96J
V0ZTz9lpEeWjcpIdaxt3nDzRO2DCbSpJHuEoVzj4vDyIU6NUV7xc6pi7QhZZ
wKvKO5cOPzZMKkBYIccGYpPf+/Xvw+Oeg3B98MwFMJmn8OEaQKBhNdjoaNdC
ecLdTb1KXHfi+V2qEu/hH1bZb7MGhBk5YftDFyd2o+fFpUQd+1T2EivspsF5
c7FK9VQHFdYIxsTy71wsssymv9exImVHEM+EbyDN0ztf+m5Lv9jzwhexMuf9
V9FUVPr3FQL/+L7rqqlAA/HqCCkipIW8+iVoKvuu/1d8/b9GU2FFBSoEex/f
sn7SW6jBdZf1//u1yo7d/Os0noem8SCXoDEHKSfABf3mYRR+suBmcquku2q+
Q4rPPpNWSo3TxOZDiWYUEosPWXOIPuWajhTFyi3Sw8kkLuQL6e1R8HpkGSDg
z6S0NMtxOO4WTFbJ4HbEpLN3X8jDR+6J5lHKxxTUSBKVACNUAf4b9fCQjuIK
Y8ihZB3VsRASNKUctFFIywLcj590EkiL42USURMfInjhsrBbaGZr8VObGYqY
nhsYgZCF2QdO3AOXYa3Ss5cEw2ai7In9PS4pTv1seEmT4IcqKHQNEh1iuKoG
UFm8sEV+1Zln8SGyI5M0EMlXeaOV5Lvnl/0cjzsbjNtwlqLBT2IFnO+REUK9
ldxHYGRJuW81XCjbMNN1Ir4fJ+JImnEf4Sk9wc96l0+Ty4oNIx4uFG8mDwfn
td4TXZ7N5J8s+WcW7pjt/dfu+Vp5385/XXbZLJgnk7CA1UhCQf59pv8+z375
lW/yrGzfv//Qe37JXugIzvXf7/TfP+Lf7Hnb/boXKuN8ZIwT2yGoXAotrtDc
oucxF9X3RrARfcdidHLISLxixmYGpKA5ixnq+Z8b9E5IsoIDnxSaLtlfGleL
Shyfz6FHVesZIaqf1iFXbAd49i0bt63E4ZxnxRyWb5FGOfDIWhyqboP6w+qu
QuiQ/rthWCpJddkJf1z4Ck6F3vCVkzOVVo+Iy5arFef++bMcpUmbs1y0cYGH
TeI0QEIhOkbtT6t+rMHkvc5lCEFaZxrH0qLkdmBVNreW5mbFngJtVWo6hncR
cg7fJEV7dvICA9VH4UhruebPDBOAa9hM8wdhAn4iGz27fHM+lkwnt6OENcID
gpci2hqdHhRKS94O6J7OUwt7MSXltvaQuYxCFHkKFJwL4Bce0oE9oMAe8tW3
PWN4SOkJHKC8DnKeiM8NMgQqCXElLWf2b6YPmDmxnjmoVYKeu1BWzO+HPU5n
dFlKuxeGyO+l4zJthHIMJtaoHOPzw+DYogvlLTyS0L1yp8tg2EtFZ84Rt8Nw
5u24z7jhtOZjBmi06COcKcORXDENq9a1/fu/PAZZcU8ZgfExOGIXUlgNFkvI
ZFT5DNbs41fXq/kVV44tPkVSVLjhD08v7ZuHFs1xbkef7H4jD+vjsaeZVgCQ
dDsBJGPcSJCipSXuq051Q6zwpFFEXUWAMrsQIKN2AY7j2Iq2ozUQCSRMr7zc
9sbzXznVAJ1cOqQ7TBU9RPOf58TNPYjkCCxSYRhfPkCiLelFxH7HXtLN3B8F
gkP28gYWrnFFz3w1bhhVIUCsmYFiSdIu6jZkEGTaSqGAtBE1NY9qM7BnvhjX
o2EiNTjXjHF7h0IJyyTrjUd5MrBOrkMk0bmQ5A+dkoJ9pmCwIkdXRX4LxChG
6baqBqZLVMymiKgAtRoo/mDK3K4Aj+Gc+CoEK5cdsumJB9w0fhOe6tt8mmkP
ViuS9Yp7p8BCW+2MiMpgZi5WiP6R/yKtiyelWf3oGMCQdVaEO4eVFf/3S3jw
H/tv74O9oX+d/unr+MGvo3sOSTF8Q+PNjhVcUnTF8L34QboVej5dP9z5xV+I
2WIJwoNsg2f6DXr5H/ybZ7OZ/1kHgUX7BUP4Q/xV2Ew847/u+6rwrSIdbjzP
v9o8n/yKef6TK/uPU4+q2I9Nxd5FRD/FRARXxWXvjjZw7Xx1BzMUECWdpSXG
+VrsE0CeZttNk28U2ejYJ3G23Vh40o2gAaupZ1lbbIup60IMWR64JJ3RZ+9q
z3F8Y4AEAqcPTjGTsIFqTFfcD2UOcHSupPT4JtbWu1U8nKxRDJw0KUkRY0bE
7BiyZTzQVnYI9CjokFnLD14XLMgIei6XdyLzSx8J+Vi1Dwxouiz6G8l3fJ8Z
S/AQMA6PbCENSvzntPPFrxzdCeeFSVaMH689M8jE/DXjdXvGm+0Z74noPdz1
oBLb30p3rGdND0CrY2qU6vX2sDF6CWZiFAfiZm4eZ97gkn4Fp/76yzeEg/55
d++X/L3Zl52uX/K6Zp9z6+qAP3uD8MmLi/P9H/r+4vyLN/wXLGvPN/slT/Cv
cgVnn/EGf+aBz3mEI5fwN8Z236mLt0e92bLMGXwATHfga8xGz8e+NWYbxYNF
b+v5zNACZebexfVuwhzz9B54i6P2kB8/Pn8FFGILT6cBG4QyW0vEs/DXRaWd
FxjypTmHTjciGpD8fmfJ/c8lS0UDY6H4nDel9TlzaaA8i9FgdmTn7hjeRLRY
GkkVEi91zqJFO8MLLxPsBU4Q6gpLn5N0NXUa8xsT97tLnJwylBvfwTItFECx
g9znBZI1a6Q73gYjifeZExZ99DTv4ZEDZt/HqSXNz3kG2ePchviGAgE40jgm
beuZKtGixZoNsBd5XYgmstNWdc1AdtuN+W5FZXNQfl5AkUIvHx7b29/1fM+x
h3rZdyZbI5/ftOLTnmVvrElHLO6jtqQaO/DLEWVv6fDdkObnhcLoecrnmCmn
/L4KfgOepSQVh8x/+0qSW4wQvsVCIvVF/Ydpeqg4wjXFGhCf5gmP3eCaeKbE
YtCEKVkhUVnRFTRaKbOLgf9uDMvK7X1YjKzYDDeIk1R38zi4ZWjvhi4JUpMx
8d+dRPqScUaZtl8paFy+QeOgu8LnQRKtUBKxCz1BtFLflbdF5Q+I1MkItkk3
SKxIqoZINk3SoTkFwbegN20FME4MIVbOqY/x+MkKi2Kt8J63lCaIibLmmq4l
p8X7XsnpzsS+qZrrh4ppt60KFylak6hAG935EHtDOAiZBloUgl562E5AW3N1
93VZTdwaGSlzq8dfcHmB9HVUDNReU+glt1VfSYwxwut3KV5/4ncSTe1F0wDK
t1quoKidSeWIKdlLX8ZsbRr7jnwPcH0DTT2wRddv0xBcrQoWQ98ssgIDAOqR
1jyhUvpeLjiPVCk5ypocYjEnrQWMW3eCCTTFor6u0B4kcGNfIRxY2xWaISKf
yRLojycJDqLIcjm8zAVoc53HMLZXSo5DLtlldOO/ZUdRDpSAW8o8QxMcD0z7
7NWbQ/6f28dJXgMpFNfXXAHuRmzlTFHRUSz/j3FU2rIjG6JsF8g4nwSY72Fu
/BXRylbxsQLoPlZDGyZZO0wNfi+5RNB83AZWH9iT8OoAew/UbalxQ9dnaxEI
BAwMUF2qnhXLurTGUKxxvHWfq7IXz15PUZ8edTKu13PfS9ChDCKYslFEvZfm
+z5oF6EiD5BwliXs8rnvUKCFBrAwxXnJf45hciBn6Hje+bKR0HCDccqjLByZ
jSxxjx4sZkpr+spnGMQk7dukfPyK5GFol+zP9ArFcvxCte40XqARgn43WTTS
8G65uFFPUjTuEXaqOrzAxjKRPLRNaPVoVxQQRrjFaXbyUMBk6wa2/UQAn46z
qi0Wii6kid1tAs0S4npg+wIP7fUowYvVZrudVEfpGWU+K7djnawQo9EAVETK
6vHUWe7U17Nei3XlCZ5FqIuzCt0UUThvx1LrMOqECsXkAXlWqnTbGebELsXQ
YgYR+9c9YFB0vyEbiHyzpmG6mF1UU2Whoh7SwSxqqs58c71WMAbfPDLsngC5
a2E+cn23lhMrm+nBlOWs5L1xMUJmPsuOH8++cWu6XcbO4jF4iYAY4t8swCI/
Cx6pR49gJJGPHxVuokWEBipisBP2JoqGXvPwe0cbqtqsOPeldJE/bJn5YtRh
Tfw7PDZ5mgOHrxsoUgg/eu50mj3Aiirpy0sT5m8uap8Z22/KmEgp9NAxT1wR
dUnDO7wvKPKqhMi5QLshGTgKt6eds3wXvqao+RQxv8qiobNo1nN3l9ZXO4ZY
0RCe762daQ8IbWada14P3IgckIDuo2tVFkmgv8C1wBbCOPkIStCzKbwRiMor
C6DPsguVBN4QEXDNEB/eZVM5k9RPou8+v3z2BvbABQlvEt0TT7jblSD7PhmM
8j5gwq7QaFUJJLTGGhBINtJpjZVUoPqn7R3yFSzqOLl9B8lp4rATucw9gmgx
Sisab43FJt0KsJwlMMxoDwvFiXK73h78EqLyoUCzDbKn7GIznOthqkWtUfb+
FjSdS1N+LTGLQ+elNi+1ZmJ8u+/xZwbZ0wRix3IPrPpTo3uQZFHgP3QJY9Qk
xRtw0c0RJhYSQUwcDPNEWusZGqXHOcmD4iXQ0KC1760iHYyMjs8K7GHHq6jh
FRpdSc8rUiaF+2tpqo7Hj4YLsKJEDt/mUd8ubipTx+WwJGNHL3N0DMbyOzPx
d75IikEAB47WBEvpMQ4Jzypnpb2KtDDKxUfn8uI5nYLL4hoWxMVzVg3opU07
DhaWlFM9yTD9VU7aaIsGhfN7twW708OkZEImltgjo7N3342DmMOq8fcMJqI3
qZht+ObGezIb0PZCFDDH2R5R1saTIaY68y9F1xFwobn0bwvLoGOIGhGmufVs
eflevQAI8ewzO57i/1l9whUdB22qa/rKxLAJ8Irj6QndfxIu4+0nU/y/vqv/
q2xjjGunVANTNOHgc7J4OSEp4pMcRpwLFi53y5VyB29QkJ4t74Xo48mmNvKA
nSgrL6/c5z8F7hdDVqmbzko2djDpXmGCFiLgKiaYlMv4vlohqz3hYhGjCzW5
IVsYRMYdirgOhMhcMLbidoY9NhADlKfeVHH2wloFFBMQ4nite07WVAP20joK
H1l2lO0MSBheAxe1W40XNXFamPxLXRAqO7jPjnPPy1MQH5EV/UJG6HXDXAf9
b8yUm4sUY39YT/nI+srHsgYAGPu+eXMA0yY1Ukxqz9udHxMs4OHHaAWOvvRF
F75oxKrf48ePwRh7Fsok9S3DtI5xa56XNghxH3ca/ezZOTs8IKaSaUn8vI/v
z/ifxCUH/bbuEwgvlu7mZHBp7rrwvqlv+td3eYx2+DzGCF6ar4PsouAfgOtI
ugCLI4PpOYZow/tTJ4XziXjzuHVcadARu1eyLbr+HHmF/Dx/Zx4Mueu50YFT
OjBp0F/qDE1RAhMYLLkQya4VZ10HJy6XbLxpcXWFw9ZIgo+4O1igTKPLpmEy
NZcgjOetEV6kcSEUf1ezU0UkIzM/1WTl/ElrWGgAUavngrgGczdV5eYCzYme
Fujr0/kuiQGvkr7heuceEODgZtFSbStpwVosOSMI+jHbmCiikHVlq8uhebk2
WONE99YiNsxJzqyaWq3GyPujRrpws14W/8evUktSBVjfIu4b/OYb6PfApWWJ
LGZUTDSaUW+qtn+LtGlF4dt+o3rW6y+p4XNtNdV5CzrUe0Dq6sDN2cvKnC83
1/5klpehM/LF9TuU79CpydlTPpYXJb16SO4oyDGJ4ZVyIEgVEJ8KtKuOX6TN
ha5ics3qy/qF8r3kCLSMJho97DURTkBc2baU+AP8cL7HMaotJC6TQ9fIY5+L
1YPyfkpr+IDd2yYCkTb+68Npdnx0lG02a7Ldcs4zh7ucRC6NYGw52MePjo6m
HIBwQXngx17NN60f12RHUMR7MQICsQZ7wfm46WNhcKzMIlrDwKahElEKNZl5
FDnI4EiPOmcPiAK8CKoL6PpOAY717dKcYSWFqHYD2zAaPVJOrc4GJM5vgXgO
jZSV0d8Z2d3VAD1Q7nCvsEGMtASY1R4oc02zOsK07GXHR6IK8Uu5ban8Juwy
aROvGNXQzkUdLrvUaLKAOLMc84BE7Wd35JV7yljzR2Pn7zFrmQtvQYYWKOhE
6P19CdyWNsAzv/GOJsfAfhaPGyQJY6itN13U45M9rly1S5qAEJ+A86luBj+V
GlLSpoIDpi5EBDnUjFOtuexlY4hwEuD4y9lrxZyWBhI7XW6lx1WD90gqielA
aOesiZM6YngGM3YLYo9PHh1lPyIlCShr3HRxJ5AYr+13dG5cfJ7lXVt7F524
kyc4cBHaB8LkssXSCNczk8BsesPWV7LKEDpreR0QLQGnpMxshKUzll3kzKYD
fqQ4SCIU2FMRZVG0IpQV20PEMJdktx5se1170dXzXLkBBYbRBpeitJtmPG+l
DsP2hU8Q0c5bomscVqvXjsXRgCvEPkshrMib7MJJ0vYy7R4BJ51woxe3vYxz
1xuLJaAT/atkMz8zHk8Qo9SPO9mpCaS4fy6kEDJ7jXXNckfgEcl11vfHEPwM
Xnwn37bV8U5x+XoAh7ZqduXdOJ9OkUHF3GIl+axPziJYvQSB81tSNpNheCox
P/wRKGSW/UWUj+PHU1Zsky2mg4EtCqsv5Ogk6glHf7Re9GYIgseyEOq3tu1U
VZQxrdmsiX3tybYEBZNd+aTcdKZSSZzsgmwz1gB5m1/Ij3G2IeuC2kuTNk6u
o3EBd8NZ4JaoJ0srVpvZUGqyWfqWmfPB3ZTL7ssNUYmuPigdo5ROuDuCBWzz
LnIOCLRJFOBzaynxbCOAL8PLDBhq5rz3Jq09tTM3ZdSOs1c/XL738AI3pNKw
sqIsKxkRGiiSvRJ0p8iCBdRZz78R7Eh7joELaDwG2womGEDRFE5IthvwDvB9
BP1DpxDvAkbvzQd1xsFLha0sGgeBzCZR5z8lrXqipKMAHsyqB/1bwo7Eh0KN
gRt+30DO5VtpChhqJJYRPIM8l+QGouGh4UcU7qckowwS1AfGc/vKRHXriGyx
Osq+ktU5+19w9uVR5cVNjmBks7teQ0rieoil0SopEvsg3ARYcDq16Cwvedf8
nH4tVD2vvTXOThYImTaO+cd9CZFJxu+Wzpmdphn/po27mZoA4Nw6NlngINft
EA5UrllDl8ot/RIAbD2yo5qOQArmPB4UqNVQU/V71vRQzUip60EbGHbBEWXw
n7/KLkUyR5dbTsO3fjESoJRsPmRXWypWry9tHwvQoImcb3jMbjTroQIjxvKu
+41cWSmSr0uzmUL1cGf+4MC1VK2whkPMpUIrKLgEg7cxAiJZsuwtg0tcP8f+
7WUteskIn41YWTTyRsBwMbGx9Yizu9QcQAwmTnVQz2OxskYYSc8oPRmmxpvp
UvWS+PQgSu5vra3HEtcaPN9Co8JW70NTjH4KcCT+obeluT1dpIDEXEC1wB12
XIT9LB6Ctst6CUzSsisA0MqnSYHYQpwW8YhEnPiF7fUijNpYGiFFcKAy/MCP
rYgX6xoKReGt8BusEiJthDOc4kxPzmt2Pq+SkzMkKKaFqy06SkX9SU2S+v5X
PWiDsnGD5qUTMR0tuBfkh3fpRx67hHG7/S6QXsSlvwM6ATpeW0MLgukwwN0A
5Dg347Wuy1DK/TyNMpKENm7QFGdmaUv1eQHPv5VxLOUcR0nOylGU0IIHKXo1
6QjEU/XtZldrhqlwQKv/UZ7mEjHrEeavNDkq74JigCictCAhLUga73Ya9CVW
jH0UmOW8vSGa+Z2aQV5nEPI1qNrC+6tkaL9pd8CaTBKjBpIJiSEDArSI25BK
+As/RY0zVceShOnWaZZ42qMoePNwVxx+yc2TuxG2jnZB2lfs0Hvs23BWqunn
D4qHpkPnPs/ak8kxNl06I2hz+6c1YAN7dA7jCK5rCqHgtXrNSIyrHhgpzbxG
Wo0GTWW3guK8gsKhxbxVjFAvz0VQht64vE7vbA19Ye4OZuMXOiUB/2fvS05S
7UynVk1bjVglFDdc1v7u21bytWlZLfJNLtkzfjDY0jAKPTFffvMk+MimLHuk
VCtEQ7UXhLIWn9moxxdZDls0PeItW7ID1XdAlMZCwaelZbSqApu0NTZnEIWi
X2jBmO8SHSdO6ynXZC53LXl0Ucc7ddxxmlPce0/2vOqZcWimqE0xNEwjnDKE
ncq23UrdhB3xPOqJ5JtADvr++hZdlnN/VagjWTPmJc0QDe7Rygh91MrKseYg
6WFxBbq8JWC2siYpTUnxlVPnnqkvJ2Tzv/YroH26q3bDeHa0d8Y9GW5PYKtl
jznFy/MTONLvkVIfNIchrojv9xhCaSHZf23Uoq105c2hXWh3s20DIpsOSxfW
KyBEFM7HS1a58hSN5qT50tAeefNfh57MUF95/Mj0lzUZm3MkAndIWp5EPqN6
fltKG5TYeWvY7iCXwaFUL5v39Moql62B6xtcADQfN0z6McCCZF3T3uhqk4bj
q7Wq1kds4FCWVZ/fJxmodXK2J7HZ7NJqIAmcdFvFX491miRpXMvvos5QISfa
Uh19I0ueg+dKCbYoeJDVlMiWOzPlcA4Zo1/xvPwyC/9Rnn6uxmMaoOOqBHak
mPsojxJj7FQlNhurvQmfZ6tR0HCcZVEsVvV2mfb4KxVHteMcU4ioP7+ShDM4
YxfqJpL+HHDx8ecOw0curWlAlB5uYf+oo9teLwt3FeQakYLjxa3ttLFASdaW
AEuxugrhdMuBy50691cMcOHZWa+5rb1mHBSY0lIBaZI0tnpNinA0B8uSNoUh
KuFgrDp1gfDCSZNEX3+YaIp0xOOXttZ7TrhBMD71pWVlRSwGHsEyEPj2Eg/t
/xW8VQ6YJxae3Oi+6MY9G8C+rfc4xv2JIjzeVxW4bt/hplU7apTFBqrr737y
jh0VfF7dAKigZc5BVno1k99jDBMqU2zj2pBIzGocaQbhwlwNqB5BiQB+lJqW
icUOB7I/9cznbXPFjoknBfLSBOTh7gSW4eHIlXWU3cxdJuzDG8fIdd5WC0VK
6TQkAwnPaltoM3ygXYad9zkf4IwmlNWrxFTpzJ53RLrEPA49utLGxQmJM4SC
6FwKvsqoZPI2tFsW2aBwqsQTekrbor6V5vOaw+U1Y2V4l9s1FzspdKIcwYpL
lrkHKadUiV6g2s9h0fd1+5JL1Re4iPamWCEhUtOxNBuiWGy1aZnPtEoa0sLR
pU3TTQcLdgqXXwJPVjtAtz6vBlEpJk8Joft6FB7kVbnSvFv2Yvr0EdKhoY+Z
HJM8Bc7V1tmmtW1xVakO7jCo0gF/zvfaBtBSH1Own/SRdA7NPn7VF+ifApAB
f2creDWHhrfj3NN7tMaREo7s40e6/OlTQBSC+Z3AXfimRb3UBYOhSQGFOGMn
b+8DiI9hCGWSBGUJNjs0kTTbBkPSUZ/yJA4N7MgvnUW0/VHsQ09oSWc4q76m
dZ1XXMMYQJgYqDwq22GvwCSr6Bk5wh4EWXQ1+H+v60K1AoX+0gRLLfHklEX8
XY6NRPtYgiOTDqHuZFqDBYGrZhtc/nTDoiBNxN2Ja0MAV3GXervLPtZFJEAn
ljE5kZ66UGqXwC6RKnBRpdWVESGeJs3KB4PEgeVMOciQBKAvLTfF+2Gdp7VM
ad1iyMpzISOZYVDnq7IV6tyTZx4ahiNE97ww1Bdf4KSWYFwSaxD2XOFZqvva
bjCd2FdGmTsK4IO8hpJrLw0BZYui3Ng+ipx0xUgqrJQb+FnxcjDW1zBzO6TB
emtB+5+Wa4lQSwZxJKZd6pVWbUojB7FLQrAH5YM+znhRqcKxc9PV0gqweTA3
PORh3FnQxRbwDsfMzMDXYnIxDVEMRzYlAjYOSoWt9qxQRwcIzFNJFJNKaEu7
e7HmRAwdLneOpgzCFGEkvlyov0JiYHUwqV3udQA2eEB42qZX1hCIYU3UnpBe
flNrtfmgso92RLBKTaX6+BVfmdrvVkSpGGlt/M1VlN0dvyKC+ptLieiQ/8IJ
i9QqzS7xgG7BNyw7N7Im83AXr9EBcYLpSdnJ2K+k16mSNMlrRAqEAYmVODXv
vZ6pBC81lHgGrUqajbF7tjYf9sSwO12onYHsZWLyn7tSJD+io3lJw+QuBFJW
jCXzzqCkfXjkGNJ8Vg0V33HiWh/dlAtiC7a6GnVvD5b641fL+AbO3NTO6qNN
U0xtFQfPRSlLaheupGyCITXjLx5cPn99ED0/dBQbxfryEgDkkvq93mxVgXkh
tx9kozfPXngwRendJ6GAID6TEgltrWEct6vp1NTX994V0wev8JISRFtUSBuL
wFc508zqEDQvioVSAXOUcWGF5Wv5QS9KDIVGrDQX1k8cXO8u//xm+v4FBma1
Ou88V/cd7sTINAdabHFJ4TLg1qWOOdqEwQkjGQstUxAeaIxapJXCKJr9J/3z
BMfKKQyrsB9fziCEoE+hB3pMA7RpkpsaoM69/zF4iwptdejTuvew/L51JE15
2ZOPJ1n5N06vCPKimcRFU6GYKtpdelUf2rc1dclXlO9c4B3lS31UDWAEd/5V
AUGTd1yoQf1jRgreF9xfy4kL/ZP7Mx29vHzXjr087CPqth4lx11894Z1Fs1o
D8UtoXQ/UHpUAhidrtB5UdUaa6ZCkr+rF/Uq1t2MmFYa1F+TYc1wT0iJRoMz
1cZpXFqwKkjLX1hofi/y+UMhovOLqwXNTEh8ktXFVrFz1OuXI6XoHUyu50Tz
3TesHD1mgmFL+u1CIaiMqEqZsfPp0lILBoNaSgA0FTrdAPa1LrxRX9URynPx
4YbOjdQlgejQF2Aii7KV2LYlvnkBEwAW2PWSAV4bjj2zoThBExHSsKtd3H81
dtEq29YpCt5SMPh3FXfSHOHRheURAiM6R76tMg+OEvhUPd8pjEyXtz8qtz3o
f2fKcuEAx4398p4UZWoKybJhvlFK8B15zEgl8I4/NesifFk2zmwv1b2vNklU
gxKEEas4rKKyXaCsh0MLlRspGsk4TDtmpSHr5OLN+z/DHL84e302cOu+ZlK+
l4YB2dmC371iMx4vYdaVVz9CQT5fkf5OAu9/5NsCZXqZFLWRiMWz37On6oLj
PvTYX//j3fmzaUEUUjentB0Fe3zFZvtP545ngrIBRnaataS+w2coNPy/zl5/
Jzgje7GYRsD0IYP//bO3558+jSchfZOB6EIcIFJ2GEGuIhHCqTua1ZIA55jb
nIl/k8etar0tTK8Wc9jcCNV2PRd7hH1MzRb1DSez7BlT0aILkECeI24ruEx4
t16W1fZD9iPnS68G/hY99Gvx5graAn2eX3GJYjj2JsU5wbCgl+WH0Og3j/oN
SU2AOGVgspMO86Vduggdej180GhJlNtNC7K4mm5K0yOynjJc9ti5oyNiBCUn
WpuHir53dMwQToW4LsAexBN92RVXeUXXT3iURP03CjDPfdCtiIYOj8gm2uu/
fIdI9FZQdNzRA3tuwigS5UqS5M0hwcm19TTfEltq6OaH4WbhSfFXUBYc3s0f
k4nRg4/+2Qcfa2j30tKlLxdIinmff6irei36o+5kpD7lSwXt6R/To2/ItP0g
2Fo3rEQsl1iZ0Q177lpafl6g3jaVRXf1xU3CHv2Ji+R0qkCWeMflvlsp60fY
iivsFnVDUgBa7rYDB8RSEMPiGjSpYHvx/vz45PGMJt409+jy8Rdet2ut0uus
zJGPcgUbkUjgh80SOSh7JjdTogUh0JUN99kzEUfztlXjrzXFbUkynyyxhGGd
Ovdb8/Se6t4I0rfpirtVxRQZxvRE+hCeGwGZI3XdWJmumvR1Ne6rWKqgorCA
3adZ1Nnmc+rDjKbwmuZmJ/9gn2IjpYMHp4naR1bUIal+9LUhWhxPTmy2yKoA
q50YipRCjyVGwfBVwn32SdJI2+G5CEEfiM+QTLzpH+yXywN5D4KeHaNhcIiB
r1eyzvIXfsk763b5/ikXD+/uorRg00GqAwzCTgHjxDr6XiKeQmQpqZj/StrS
7UJF31WD1CqULy1R2h2CYQPo9c92D9PazNJEQEzLyFEhGo/Vv2WDh2Phz58Q
5hJonQ2JBSknilLis1ivaSXEsdKTQaOPH5f8ZZKv/D5h430NwYhxBpaM0D5i
H8yAV/l2Wczy8jR7ARHDTOdFdQ23Fi8ny12UVbGLgo7uiBFe19yP607SCyds
7NZc0cQG/lZ9hj9ccoBjtZLQ9/2GLDNcGXMmLqlgLOHxemzkc4MzJvbFSXKn
IRA1ejV9OZYuIqOn9OMs3G69Hk4T53zAuR8JYI+0A++j3vN7Xu5tkokQGOoZ
DawoaonZa4eJz6BcCYP7I2d+yy0m7k+zNxfn38uN5/TTmJTFZnnHjsV+DIL/
JkKMfUt5s7jhd7Jlzcqvdngaekz4pveA8ITj4zRjnSsFVrCwINSoAIri8QsN
hZDjVPI+lYL6Mm124PN82KLa3E83cEihFGoc+AWAgr2n53Mw7dAdg7LGEvBv
HLwajfUS5t/3T6uz5XQXZgeKFLMR4xdeccyKX3Dx3as3zJq4ln6SvXnznn9B
UZSnJkVW1bKw016S7kjrqY1Uxv2CJy0axUS0XMhGyfrm0UO3/1O+GbCGiasM
vibYejeFpLPqCmgZqRWvSQkanUup/owLRU8z5Pb2Skk10VDv7qzTp9SM0qJJ
STqGPe4vDrGplaLm3SSwhpIqLiJoxO5/X31TfOjGIcndqr0PacDHVi0VV18z
wQzgyn7HCgmXrmimkjJdLdpjgDzJgzt+rIVPe87gjiCHqOoKZ5OdvfsuoHgY
rtSTCARlPBFp2pZt5116iBTvAD2JLSGmRZn7hIcZvTEw7cg0M359GplZ8esQ
VdMi44o3j+XqeBIsGdTQx2bFP2DV7Hm7iHF9JxEXswT+jgUTK5QaleLAR/jT
x7dR3U07e/HixQuSeMvsydHJ7Pjt9OTo5IQWdHRWVcWH7L1mzq1ydN95Mns8
ezQ7mZ1M9MfH8sOT2SPcxz8fH49jyEwN560594hJSqVsRuOWOckqIAjCY5Fh
4G3h1/ltWLpW/VPLreRMxM1hWcO6nw31VQCDMQz65kbcAXQUtiiqqrpIURCs
Hp9xOA2gWnFEMos4dRRDN1XVYkMTdQBGggEedHE7+RYzpk5YkGZiGrIpuXtV
XGjQSXxHKVQgKobaPjiSKqV7lH3w9bbYkj2EKkEoA8L/FXZ2NM56nRdJNyVi
QmyIBXzI3AyFf1Jt6rI47ye0lGxJw9nVqRE5WR8E2i0SQb+TX7TFJETSB6jU
84Yx4ifWfLKNy08Upzw4IfjldODiQqWozcjvgghcsCGcwa8QiUE+WZwXEJfQ
WTCFJMt0OgW6CLsNIr7XO9V9U/Wr5GYrZ9iE/YjOVSD6cD2G/vCeegNK8326
okcXEogX+WHr4MKdIakpESdwIIVWX59S6GOB85C2HihMr3J+1/Hp8Q5fTbxu
va+IR3XF80WggZObJCwOh3nFTWBfswuc3sFHSQPsjPE2L1cKcRWgnPo4Lnra
h5AdlVE1WVx5eFOIRTrwfiC3ZO9D01hrGQ/JFuEJ91J3LBgZo0ztgJgxP1y0
+0A4rqJkPznL6KQo5kb0VS7GT25cw5bdo+P6VvdF6/uUCJ6UZJ4cvNmS6XFR
aYfC77eIB0JrNmx1qElQi+QRPvZLx5q13iJUEsp4YsTlZYoII9gKflATFwV/
S2hShUSCpNcOtlyDgqw0MbYQ6d1+skMIn+sA5d4SO4OfP9oPWAb7no7MAUaI
1QKEyDQIIcNOqpjojVgH3/vNUCt5Qzh0geKHFXqJLorwYRmaUGCwcJNQuHeB
f8+tMbLfRmrM/3UyxmedckMUruyoOJLoCT+t+YzBNzsqtQfH2PkHPByGZPFU
yxiQbvAGRbYJnmCkX7borWc4u4qAOGZUzyQlO8JvZCAN3DzSpBVAb4G/WSfK
bSgpSAPRI8uqGUfV6TpEGjD3HXaD1RtLCJ+tpJt9DLxVXCK/rabVRYeWO0hX
2s7erDnPJMooJgiRSTzwvi1b78X2WYcBl0iYL447nyjmsPYUcV6p0gq5cJZh
7u8RLHAbMKnAmvtngX3tZAt9LCkXMp2Svh+dFOkPDrAApdr9PNHB/YTySz3t
llzFqxz4K608/2HUFkUY3bnI44fjqIhdtX4xaRl8hDURlPWgKEyRp8CxpF1I
2WoRvtGVpI8ZlU9DcQYazewu8uKCw9316fJ+HHU5VKHXt4o6HEXUOZiSM/qJ
fxvHpOAzwSLfKD2PG7kRmBVd9hUvJT7cp+1mrduF14SiV/oqTw3vcR5WBxas
BquHvxCAJO1B3P8sgMtZciz7KHTAQUOgMQR9M00C3uV5YJv8oosa5eb71nke
QeRiuW/qVVT4j3h/r7ZzxAGyoBDy2bBtj2pyeKYJ/nSo0bShEhORzJXfRGP7
jSzSb/wQfxMzaUEYcNHqBNCAnXupmQ9lVdW3XlXa3cJyMohcs9egVO8Ez91J
0SIaRRr0zMeP786fnZwcH/HRKn0nAQZd9MDlJNpuCm1GwlokCqw0ic50e/gT
ds3A91+zPRnUhDvsAbPdEZbW122O0x3IfNq4kgq0lyraIQGejHL8ekW4gA/U
nhmS956xhrJS2ik17yjAUMQZ6kkUQ6sKkpM2sdSXgJLrquKalGroJGKAhsau
Ya96qBetFXiBZyJ5wAJFBh2w6yiYjy7oVXK0Qv8mYX3IoPXwx777RLoexq3k
EXSmiEs+Qo9GzVxg9+fVilUgetdB0Nck+quMWjfQ7TzyKQ9XvgddZHfD1lGk
0YxD87EIylsRdaJ+kpVCZfeyVQc8Ntp2a7YkvNTC2fis72BShfKK3QzZYEIU
1SYtudAe9CoAd+/tHTqvyxhuJNhubS+j48DKM448l9Lu4SZWDBDTHwlWYTOC
s18aj+a4Qie9fyCY9QiJaoEnAxCBtrcmugrAipF2wcphUIUCeAULx1kvvdjD
lEUoDV9kMwZC5VRqDniFno6Li3Or5gb6c6pcRoz5Jl+6+LIqkPwzIw1EDC1l
UqwvTPiuidyrtcwuYlQ6Q2M34gm2vCaZllV2xbl5SL0VvYH5sZ1YMSgMudBT
NYyNodqhqsNTru3g8cnW4Ce1YEIKIFM2dhguPwC1LfJNa33hoTao5r7De7sM
dUzWjuQ30YYAekX6/EUtz8FOUH/QV/p9SxPwL8291BdEfvWQ+BRkSYx4bimL
Bm6jeqLtJdZ0p7IXWVt8FoQIfa9nkeyTLM4J3AULkEqRqDchCm6kIHqx2K51
jbEccjJK7XuUt4qoGbd0iupzUyi4xMemmr3r3TJcKKLvfo/ZHsmi2tOXSUup
j1Au6VdzgMNyItdwRmM5eD3Auh1DCO0PtDpAvVihSVFfMfbohShvlUKnpdTk
oPv6PrXJdMc2QOawZqGqZ7Da8aJQLNbk8HGgSnf3q4XdPgOnLVvpsRRcambR
AUUBncTFlrbrsQfMmhE/n2TiEX8lDZTbmOKeO+nMrSwgakVb+YgP/eVttuO2
fuBPkkkj8867i5Ie9r7F+nO8/q3B/6oW8fytDAlkXHaTgJXQ7NPEpqJx8N4K
JxVWypxUXEijlj7PbT1EcEV9xMJUDdYngi+y8iiyrwNTVA00N6twBNocfIBI
flXEY3eqUxTLQ/MT/FwsE7OtRLy8/ldu8Rs+v6fFb++GYYtf2uDsnvug72vx
u+OGXS1+ex8atvgd3vCv2+J3f4/fzzzwK3v8Tm/aqbbxtW6/iBNcigtGTp3Q
MvMxtPq9qAaMhZhJaeVuvTCMyJMNusMkvb76UgMjvecgtVdiTEKDESadsGVI
YBJxK6WeBSB3vZ0JQh3yHnaMEFknnH80daOfs2l2Px7P/Cz3utJT3gN3oTgl
FCAzHqgWbIIrpUNMEfRhUd3HJqi13PxeAXcrUtPREcb6/Ub+jwUbidYz7O7m
3rvIYDnFrVcwK4dZncKLYLJQfb7s8lK+KUyV6ODQc7kwGsgrjMeVqgB7AByF
J1cfs1WM52LCd7Q4AGrBnrNPogbrD9hjgbpSb2UqDyLsHy32dzLsV9koyDHU
G/R2PHiw0Lc+gjKCIgWkmHs3DwV1OfKIz2n1thVAFScZ9L/iQw5EhC8Ya56W
vaPoudPRp2hK4qz3N721OYr6nq+4L1jS585FzjhWfDlQzVeevz2lEf4+G33I
vrbpT7O//ptmDfz1D2Pg/vnecc5aarxP3cyxGd0L16rZZaamp5XQCTSMH0XV
aXS197gnMDeY/qz/6CjCJ1rdI3C8ajMXx1GFi/jeAm1MJcoZXmk1ZBTmBDxV
MksRuICFTi5s8h8ndXk1zj4qE+54kWnB6cJMofHkT4eHKflhesoE5VGYS+mj
8id6VJrOcI5T1MnZnsNd6XP4Ez1nKTSDZxt9uCyvMvuoBsRn+Bs9vMOB6e1n
PG5zzukF+fLvf2sWt39jah/RwRkRZQ3+++t/HP/nbmnF80iHMJO+6L/qdsnX
GyeDWtKgaGRfZ379R4bl4x0a494xdcNXy3uWO66s8w/HdI3/+et/EBFAn7Tp
xQRCj0+yn0Z8S6m3qRkvI/7kBTGC3vqoieJ3OyLxIWfBUyikciABg+pRkZg3
kEqc6ORdYbuwTn6nq1UOuxzM4skbGMeymEqrg7QA3HDkmZL0VRENVUPhzslS
Nh46jz0fQq48fLFtGjECl0P7sowzMtIu5Aj8JfWal+/+yGm0786fPfnm0cNP
nxAggfc6KlFEJQC3WNMGknBl9Fq+yzu+ffKYPd0++hd32hXfHvL4tHDQEbtl
P1IhSHPZAwCLp9loGp2KXn6pyeUPZsdjErUB8yruCBethTZHpNV8Q7tv/HW/
/yJPUjh8ynIoMId/y/kUr949PT1pXyzavAVRv2BtRJRYaql97LNSOHBu5fIa
enQbnV2ZaC9IWaVd1saISrDSDK0LbMz3pUEvZsm8wFMNC8ieo679TYqRQJtX
bxgYwcdGtL+YC55qUUA2BvLG5HQoXfhA1b6LEm0TmCY0pDZRkVSXlHM22lYh
vxw1ZpIc6iFNsSbFh2Kx1eQN5rkpR8uzkQdSFQ/o2FytkRqBW2JFXEOd87Lr
aTPQvTXVbrfM6DSeP3OJmBjH3WV0tZEl1W7nlr7OjFwH19esGBO6mOyMQybJ
YTThxJpPRHDErjg6qGBsCwBEdRa5WqmNk3YCYlWei2qB/V+EHaNF7wkFmwEc
dT1ARlsnAyaD28mCpehwTiQ65SfHvcip9ekQnUaK5DDhaEXYyz1oca5LPUj5
89yq1xiV7tyU3qfiTAYkTfDkcBdRCyE5dIYesNQkhSUwT4K7ZGKmHxlO23Vo
jM5sVqMlWr4TFTJrBMu6h5bBeJwumP0tkOtlRkua1a2qqGJSaXsX/oJb1fWP
240cTGR3CPqBlJSx4zHq1qcLVuQdesUrjr1XftlKIAsIuNpp1LrvoIJPaodq
ELFAsRwVFCs9QM6dGY2aliDn9pplf2xFf6aLFUlojb6oSenbpkuwljdGTTnp
xmVirs8QXJhmW0hLS9CmFFUQ2VQeAH43r1DHFn/lb91ApZxIlsxEyzMipVvu
p1tMqyWZlNwbaZH6IzH4bVMheDe1eo/fkiw91FScaXb+w/PvXvQUNBrSlIYk
HhVV0F4MEylRk5Vs1MEneBOQeq+QxEhwYmm1YRnqYxNIckRzIQ7PmNfGUAMR
HhkJVAwt/JiNUgZOgRiWXGzjdQMGLvnaBjvDcqLHjTUUxyKoMsayLPaCs0ft
glwv/ruDJXs3Di/1RKnKWoOhcloSK4Z6fXDGMqk1mq+ym4Im+14VLAppTZNL
Xu/EFnApGC9+3LbwQgXIjYQTwPd1C+FcjVYXO1JqFdUfLEVnbB1rUDYurUN6
b7ZuOdw9D7pHxHe0SNMEarZgoOkiwZHxjanUnQR4hriVs4WHeE2le1aoxnZR
npQPwguQfxyayFvNPONWsIkdIjCPPXYUYNuEtpalQkhuN7VCPW96WsVEeSo7
lSLnXtzUFZrieRQwt8QyH/X4LLmHVrLLYsVdtKRNBkobbW4iJrGP0t9Dyr94
inKK2243Ip/lu4WZR02xpEYZ3S8ZWI7BBWeR18F7OPpeB7uQeBwQRxgdhz8Q
P+SZ/z6DQun/eCWWLF/6OjsWr4KeZeHj1hNEkez9cz/RY69/ePlSfqXHFHnu
S8/VjdclBIPhg1crZjM6td4A+RhxZ3sKlmPF/7TJ9ei99uLI5vnY8xbQYEm5
XGqOgxQ6WS7DTsvf/uM577TYe/fRKEc/zTTHrv91dULodfYkyKqLGjvyAxjv
GoG8m575N965XS/3W7rcdQkzGFz45D73e/xb/DOP5OonGgMWFN0JE8+V/Efc
lr5pNHr1ExxjyR3BM8cry6Q4oYfSe+a0uz/yd6JyCYBucXKnbGE/R2XnqO2n
T7tcLDpIk+BvdjhW/CFkyd1zPEpyTWt8JobHj3rbBFXPnByc8GR28iDJpvO5
dQPiZDpiZI1IPJWVS/nu2Adtl9JZFz5k7UGlnOxqy68MKQeQO3dlW8S5g4YN
swMCW9zCEgonZdmXHpp3UjMxcObaWOhbiaE5yRlWtmHYFj61UUVNqd05ERIq
SbO47CV1wdduzhOVXhg8EvL3shjPjZFCb44ibzPV0gZyu/ZZbTpVzvotq0qn
67tisKEQkrGjKcZ76HYHAiJAf747hKNjjV/zVX3OE6JA3in2/cW5xKbOLi+e
HZ6/+e5sWOHBWyA+H9/eOTh6Wl5vdF6MsHNikd7dpEoM+1AsT8scNXA8TBf3
ixXr/Q3nNXiYoyx6rZEIKFMaGFryphxCcEfO5xG5yhB+lbiEFjflhlWnfBKH
z2mQ2ftnZ69arkb/MRac4vTvS80hw/HM9CdsiUrFad/v6/8jbhR5rsF0B47g
qb5M7YYTeejEHhrEAyK+7XvUYJSeI/Z4lrCVX+MTlmWAQziOheyJgkjexivT
vhLbD/a/b2qS9pAaBNQ6gbzaAjJkWG6n3gLV6nxHn8j94rplwncGfg0sdL+N
+CI0TODAqfDGQeYJn/BtZMP+pJj/OZyDBrIZ3oA7JSHTyaPsAIhQi/C8bnV/
SEMWqsYCB9gkSqgJReo3QLs1b0xzUkne+Ji4LVWwkk1xjaN+8fGFy4/3gglQ
6qsTMvr0ySJ8am5abRqaI/G3dGJl64vR1RDwcHMms5zutyGFLUPr7+HmKw9B
53pvA3iK6HlpGq4NQJqqjWBozOBNkkmjTaBiH8TuPlD9Thm0nPmPRotJPrVL
MgQl+IAB994Yp97yHHvBhnlU4yUuObbXuT7Ivhq7VXxGr6yBTdVF35NmK23W
M48Nx1lMTs4E9AdW9lMjnaHCcVAi+ub9mz1ILSG9MN90hpbuWzMn/lZ21L9/
42IvJZQRtbTNYcUf9haiPBOSc7Vvt8SELNEsYhprTYuX5taCjOARtqBIwLsW
3pl0/2AcC3QxoJmd8x/IGPwu1OyB1uPouuWpTSO4vRiQ39A76W8rOroxyRpw
vbWGdsPu5JaRgUPeGRQlsVceveBel3GTt0rAJHMHwH5plxzVxCnKfmeUm54i
C4Vc17Xv9Ma+YSuWTetboyMcpa+3XbFRSFBFwJEqwZnUI3PBteKwhixFbJc4
OvoA6gJU3HJupnYDLUW2i09szgqx9ddtvd6VwJMSN12FKAJxRo+4dcc7YHh0
GNTHj5f4xZRCiI45WnI0NdNAXPAZl9H23b8eJiejp5baPnzPvYMqTzmqq/pu
Cgwk1uJkbqxRlYvss3k2veJaK75j7PbV1iNoGez+HZLtkzJWUyqdFfCFoU/C
oIzsSUPf+kJ/3wYyqn3mhNmpJsxGiOQREi0HMQ7vSCQXmCAgWU1VwGAMXd1J
YaV19+12kR9/U7puMuPfubnBiQ9C9+QzsU1mQH2GpszjRCjpUcUA3tKaOMED
jhC4BGyDc3CkLYlATiqrNMhUXSbFl5zelaxJR9jtu/nsKF+RfMHMuc91E/qt
w/EWhzctK8kcdoFwdVJZMinHk5J8HyCzmpdMpSkcJViseMShKbBunIS5Jk6E
6xrRyhjVU5oeeC5denDeiXCqKVcdFwpt7TypeRROVDkZEvGwv4Pv2imIdK2U
8SNwKzaRN0UMkoTu6qReO3kVtjVwlZiIXOAQHz9K4bvkJ6j5uJwgYaqWVjQZ
lqkxdEPfc6R22rLA4CA48mYsLzheb1kOL6NSB0/m/Bwp2daXfIowJFjcdF8v
2B84RhZy06wMndsJFB8itq7TgtgCd1jH9SCghDyz1htC4+A5btAujb65IorQ
D3qggDtpJivIDwZXKNEETi1U8eVxTmIkZb/mEgVkxcGjfMVArL1GIWhqryQd
z1ZD/q1MWxkiy6IG3KnTCRfT6wbx1omzrYhR9r2oQkcPidcAANZDBch5wJeC
j15aueDbd9r8E+6YSBtM6p0j3hsEyKFlxPkSaFHee6JaIGaqPULIZ9rzGt2L
o2oOLwyRSB86oQcpodmBmk4ajh7P8kcmIXFyc6MoYk5dQEw4VS8KUJODgiJp
HTmzjQ5lRV7V8s2QWlUdrNMyr2V4gXqLLA8yKGVx0qxJtmXB0G8KpGjdj5yC
LXfaWbGHCYwsWrTLgcERg9oDBr26j1AStLae61Mk20EhQ/WtbdACqzrJmlFB
vGbGZV/w0Ba6SqPSOsvXkmmyNxfZJ/k6TWUa6ydTjnojoCTha1oemmpD9wim
SuFCsELEMGUDKhrF6PjoKPtuvlHFlKHhmrHmISuspvMgwD3FIbstG3TgA8Lt
UqoEbd205Ynoeqw6lHFrzCT7oFfcPgzjxQ5OdjA1ZQG2G4XSJtnF92evpiCb
tNsWia0tEeK9EZzz/Ki4VRuGgyHRcUyPoT97AwVy5t5EtcBiAmlbYg+0bmr/
jTSzhRcshZR0L5jNl4PeUAHTY7dCIOibAPWX+iQnsIEmLQ2MBxDiXkILj7K9
PzyWzR8PXJDOMJwFf3vLPgbpd4NAK1OctSzUUqeilS3WJJBapWwxcRGeinTA
8RG+qbZoTdm8iHJD2Y/VMSdKTODjGM0uTew0GB/0TrStovmv2FOpH3ARFhpM
n/8qTOGvole1/CptxlJ6yCXhBdJmD9rnih0hps5o0hte/HdrGoPow8ePF9Pn
sxQlGsq7rzUD/N5EE1NQCBfE9DacMd9xRwxA8WYxi1jhEeDmQAqvojroeNUl
YSToE4rAYt1ywadC6qBBm5hZZ5BQvBDEPsjgazer0spL7mp/Y+v59K+cOsto
F5ewS7EBSYkDDtmypA9Z5wdeb4tbycTOH3pVK6qLpJQyRzUYPdM5xS1jxOLc
Dz5b0wCFRfS1hynEZmfkZZPV7Y3RpdkPIEi/djfvrhRrizKyWOW0A4ADZfRM
KAEtOna1gnIX2gbAnhKnAydK8ujTdoJADNLiPTXzQHaCKClVkU1B0uVKe5kC
O5dnG2H1yomehGsp/qVLDP+R1CffMEB67wALgOzOsz0RWBExpcZqzq3rW1t3
lZaqWUVj63OPmXsWlbzXK5ZIMtmpSCQjUWu+0BQoe+aPoB2btbdzIn01LRZ3
8afRiW7Qolj6inChbtz/Q91fmpjQFwZeQIpeIeEqKBoRlMEI+0of5eKn32YP
Z8fsLcz+SAT+nBNtOL3kXHzFTOQvVSt+yro4h11wKPVELO0Bn8KEwDGR0TP6
wzW7iB7MkifsxitUJyDID+KUFI0OULeFbw1rPQgKbdyVceMuDPuRDLuslyV3
X+CBJexTL7FjsmVRv8yxTCtrKJElDSUQQjUUpSzyXlh3PzAsYbn0RNI4zhy5
UR9tzjXhcwQgRux5AMYJ82hjootChpyFrHM8mWW+q9l3ZBURGTQ623NtYDUJ
0hILqSmm1+Hu0OcleGHAn9HV4Nexk2zERSYg0KOHYzPbvOKigJNAlpftRDo/
8paWQV2nlTYAXO9xEchh9XfSexSYNljtnH2EBrKr8Sn/MrjBQ+lNQsaCB/FN
AurAolbMpCjJCmPDsMeJjsn44QGeNlYpe9QMsEypcx/W5qPzA2/mA95M9oPj
KLUxyXoAMjF4LSX3HRKRnqoV/Lkz4TLv8uP3KOBq26b+XXvvy+Jq8FpxZAAE
k9ikAmvZMgbcAzqwZXFbJOQeVF6uwhCVCtwiE89sPUc3tSWOmarvd2WjMWfg
O/pU6ST3q/AdwDJTP0aClckxA83B6qflxghtybqqGykDmM+HjqULZyrwGpO+
K/1cJJV5fj+wu1Qw6VY+zC7VnPyed91vpe84LZgHLVMpf8RCJ6AuvIObNJAd
jWOgvLIsYs0xJo8DxsnNsmfYUTnf3CbLmJz3kyyU6R6k9SuygHpRpaPoI6Kv
quYZDDhmJtyCgsh4Kp2jJtJqHo8qh6H5lILu7PoFs+IYUqCWBGzuLARX+85y
JhWHWJ8eywTMmHPcY6RjbQufK86jiy7t0Ik5StQwgrpK7KG4BYU2fa97gsLv
/m+VJ5q/2kQBAA==

-->

</rfc>

