| Internet-Draft | FANN Use Cases | November 2026 |
| Han, et al. | Expires 19 May 2027 | [Page] |
This document reviews the use cases for Fast Network Notifications (FANN) that have been described in a number of individual drafts. Use cases are grouped by deployment scenario (data center network, data center interconnect and wide area network) aligned with the deployment scenarios of the FANN framework. Each use case is summarized briefly with a reference to the document that describes it in detail. This is intended to help identify a representative set of use cases from which the key FANN requirements can be derived.¶
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.¶
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.¶
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."¶
This Internet-Draft will expire on 5 May 2027.¶
Copyright (c) 2026 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Revised BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Revised BSD License.¶
The FANN problem statement [I-D.ietf-fann-problem-statement] describes the gap between how quickly modern forwarding hardware can detect an adverse network condition and how quickly remote nodes that could act on that condition learn about it. The FANN framework [I-D.song-fann-framework] describes the functional roles, information model and delivery and scoping models for closing that gap.¶
A number of individual drafts describe situations in which fast network notifications would be useful. These drafts cover different parts of the network and sometimes describe the same situation from different perspectives. This document brings these use cases together so the working group can:¶
This document summarizes each use case briefly. This document does not specify any protocol or mechanism and the actions taken by notification consumers are described only to explain why the notification is needed.¶
This document uses the terms Notification Originator, Notification Consumer, Notification Relay and Notification Domain as defined in [I-D.song-fann-framework]. In addition:¶
Use cases are grouped by deployment scenario: data center network (DCN), data center interconnect (DCI) and wide area network (WAN). These groupings align with the deployment scenarios in Section 3.7 of [I-D.song-fann-framework]. Some use cases apply to more than one scenario. Each use case describes the event, the information the notification carries, who consumes it and the action taken.¶
AI training and inference fabrics are typically multi-tier folded Clos topologies, often running BGP only. Path diversity exists only in the upward direction, so a failure on a downward link (for example spine to leaf) cannot be bypassed by the node adjacent to it. Convergence targets for these fabrics are moving toward 100 microseconds, well below what CPU-mediated IP-FRR activation provides.¶
The node that detects the failure in hardware originates a notification that is consumed, in the forwarding plane, by an upstream node that may be several hops away. That node activates a pre-installed backup path, avoiding the hairpinning that local TI-LFA repair can cause.¶
References: [I-D.clad-rtgwg-ipfrr-aiml], [I-D.clad-rtgwg-efficient-remote-protection], [I-D.geng-fantel-fantel-requirements].¶
Not every event is a complete failure. A member of a link bundle can fail and reduce capacity, a link can degrade, or a link can become congested. If the notification carries a quantitative severity rather than up/down state, upstream nodes can adjust ECMP weights in proportion to remaining capacity or split traffic between the primary path and a repair path. This avoids withdrawing a usable resource entirely and avoids abrupt load shifts.¶
References: [I-D.clad-rtgwg-ipfrr-aiml], [I-D.clad-rtgwg-efficient-remote-protection], [I-D.geng-fantel-fantel-requirements].¶
In a multi-plane AI fabric, a high-capacity NIC is split into several interfaces, each attached to an independent plane. A NIC-to-leaf link failure, a leaf-to-spine link failure or a degradation within a plane can make that plane unusable for some destinations. The node that detects the problem is a leaf or spine, but the node that must act is the source NIC.¶
The notification identifies the affected plane and distinguishes between plane unreachable, plane degraded, and destination unreachable within a plane. The source NIC removes the plane from its candidate set for the affected destinations and moves traffic to the remaining planes. The notification must not leak across planes and its delivery path should not share fate with the failure.¶
Reference: [I-D.lqg-fann-multi-plane-protection].¶
Collective operations such as AllReduce and AllGather are sensitive to transient congestion, gradual degradation (jitter, intermittent loss, bandwidth fluctuation), and node or path failures, even without packet loss. Notifying communication libraries, runtimes and schedulers of these events, with context about the affected communication, lets them adapt communication patterns, coordinate recovery with network recovery or change workload placement before job completion time suffers.¶
Reference: [I-D.sang-fann-fast-network-event-notification].¶
AI workloads increasingly span multiple data centers. Common AI DCI scenarios include coordinated training across AIDCs, distributed inference with model stages or KV cache in different DCs and AI fabrics spanning multiple domains and vendors. In each, interconnect traffic is sensitive to congestion and loss and the node that detects a problem is often far from the node that can act on it. The following subsections describe fast notification use cases for these scenarios.¶
Lossless transport is extending from DCNs across WAN interconnects. PFC is unsuitable in the WAN because of head-of-line blocking, deadlock and congestion spreading, so the WAN uses fine-grained flow control at flow, tenant or slice granularity. The DCI edge nodes must translate between the two.¶
For traffic flowing from a source DC across the WAN to a destination DC, backpressure propagates upstream along that same direction and requires interworking at both DCI edges. Between the destination DC and the WAN, congestion inside the destination DC produces a PFC pause toward the egress edge. When the edge buffer for that tenant reaches a threshold, the egress edge originates an FGFC backpressure notification carrying the tenant, priority, slice and pause time toward upstream WAN nodes or the ingress edge. Between the WAN and the source DC, the ingress edge receives an FGFC notification, either propagated from the egress edge or originated by a congested WAN node, and translates it into PFC toward the source DC.¶
The draft describes both a single-hop mode, in which the edges are directly connected or intermediate nodes do not take part, and a multi-hop mode, in which intermediate WAN nodes relay backpressure.¶
Reference: [I-D.han-fann-codeployment-pfc-fgfc].¶
When AI workloads span multiple DCs, congestion control feedback loops stretch from microseconds to milliseconds. Receiver-based ECN feedback takes a full WAN round trip, and sending feedback directly from the congestion point to the source still takes about half a round trip and requires WAN nodes to understand transport state.¶
Two cases are described. When congestion is in the WAN, the congested node sends a fine-grained flow control notification to its upstream neighbor, which relays it hop by hop toward the ingress DC gateway. WAN buffers absorb the burst while the source slows down. When congestion is inside the egress DC, PFC protects the fabric locally and the egress gateway converts the condition into a WAN flow control notification that propagates back to the ingress gateway, which may in turn notify the source. Notifications are directed along the reverse data path, not flooded, and are expected to be carried in IP.¶
Reference: [I-D.dong-fann-fc-cc-coordinate].¶
Several scale-out fabrics may be interconnected across a campus, metro, or wide area, running IS-IS or BGP over full-mesh or irregular topologies. The remote protection and quality-aware load sharing described in Section 4.1.1 and Section 4.1.2 also apply here, but the distance between originator and consumer is larger.¶
For coordinated training across AIDCs, fast failure, congestion and status notifications also allow upstream rate reduction, dynamic load balancing and fast failover of parameter synchronization traffic, reducing the risk of a training rollback.¶
References: [I-D.clad-rtgwg-ipfrr-aiml], [I-D.clad-rtgwg-efficient-remote-protection], [I-D.hhz-fantel-sar-wan].¶
AI infrastructure may span multiple operational domains and vendors, for example distributed sites connected over a WAN to central compute pools. Events in one domain need to be represented consistently so they can be interpreted in another.¶
Reference: [I-D.sang-fann-fast-network-event-notification].¶
Large training data sets are moved from customer sites to AIDC storage over the WAN, often over fixed leased capacity. Fast network status and failure notifications let a controller allocate bandwidth on demand, adjust TE and load balancing policy and fail over quickly when a failure would otherwise cause loss.¶
References: [I-D.hhz-fantel-sar-wan], [I-D.hs-rtgwg-wan-lossless-uc].¶
Training jobs may read sample data directly from remote storage using RDMA over distances of hundreds of kilometers. RDMA tolerates loss poorly, so fast congestion and failure notification to adjacent or affected nodes is needed to support flow-based congestion control and lossless transport.¶
Reference: [I-D.hhz-fantel-sar-wan].¶
In split learning, part of a model runs on enterprise premises and part in an AIDC, exchanging activations and gradients. In collaborative inference, prefill may run locally and decode in the AIDC, with KV cache transfer across the WAN. Both are latency sensitive and share the interconnect with other tenants. Notifications need tenant, task, or flow attribution so that only the responsible traffic is acted on, together with congestion and failure information for TE and load balancing.¶
References: [I-D.hhz-fantel-sar-wan], [I-D.song-fann-framework].¶
This document has no IANA actions.¶
The editors thank the authors of the drafts referenced in this document, whose use cases are summarized here.¶