Fast Network Notifications Problem Statement

Modern network applications, ranging from AI training to large-scale cloud services, require lossless and adaptive networks to ensure reliable, congestion-free data transfer within or across multiple data centers. These workloads demand high throughput, low latency, and minimal packet loss across dynamically shifting traffic patterns. To meet these requirements, networks employ mechanisms such as traffic engineering (TE), load balancing, flow control, and protection. However, existing solutions often face limitations in responsiveness, coverage, and operational complexity, particularly in high-speed, large-scale environments. This document summarizes the limitations of existing mechanisms that prevent rapid notification and action to critical network events, including link or node failures and congestion. It also identifies the need for fast network notification which is critical for enabling fast reaction. In the context of this draft, fast does not imply a single, rigid numerical time threshold. Instead, it characterizes a class of mechanisms to minimize the delivery time so that the end-to-end latency of the notification is on the order of sub-milliseconds or milliseconds, depending on the operational objective and the range of the network domain. This document describes why the IETF may need a new set of fast network notification related solutions to support these use cases. provides a gap analysis of existing solutions and where they are deficient in supporting high demand services. This document primarily focuses on describing the problem space.

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119.

FRR: Fast Re-Route ECN: Explicit Congestion Notification BFD: Bidirectional Forwarding Detection IOAM: In-situ Operations, Administration, and Maintenance

Current network traffic manipulation mechanisms such as TE, load balancing, flow control, and protection, has deficiencies in providing the low-latency, high-granularity responsiveness needed in modern, dynamic networks, at least in part due to the lack of dynamic network state information. This results in suboptimal performance, low reliability and delayed recovery. Fast network notifications is proposed as a set of solutions to address this by enabling real-time, lightweight notifications that enhance the responsiveness for traffic engineering, congestion mitigation, and failure protection. There is a demonstrable need for a standardized framework in IETF to define these fast network notification mechanisms, requirements and integration strategies. The following describes a summary of the limitations of existing notification mechanisms: Slow Reaction: Existing control protocols (e.g., routing protocol, etc.) may be used for dissemination of dynamic network state information, while they usually rely on control plane based hop-by-hop distribution, which causes delay when the recipient is multiple hops away. With modern high-throughput environments (AI/ML clusters, multi-DC WANs), this delay is often prohibitive. Explicit Congestion Notification (ECN) needs congestion signals to be sent back to the sender, which introduces Round-Trip-Time (RTT) delay and can be slow if the source node is far away, and it relies on the source node to take action in the transport layer. What is needed is a lightweight signaling method that can provide real-time alerts (e.g., at the level of sub-milliseconds or milliseconds) on failures, congestion, or threshold breaches, enabling immediate actions (e.g., in ms to 10s ms ranges) in the network layer. Coarse-Grained Signals: Classic ECN and similar mechanisms only provide binary or threshold-based feedback. The lack of granularity prevents rapid, fine-tuned adjustments, which can lead to either overreaction and underutilization of the available capacity , or underreaction that fails to alleviate the congestion. What would be useful is a set of notifications that aren't just "on-off" state reports but can also convey more information like congestion level/utilization information, latency spikes, queue buildup or flow characteristics, so that it can trigger immediate and precise responses like rerouting, rate adjustment, or protection switching for specific flows. Local-Only Decision Making: Current load-balancing, flow-control and FRR techniques often act on local information and fail to capture downstream or cross-domain network conditions, limiting their effectiveness and can lead to suboptimal decisions. For example, the Point of Local Repair (PLR) executing FRR makes its decision based on its local view of the topology and network status. It may switch traffic to a backup path and cause cascading congestion on that path, as it lacks visibility into the state of the entire backup path. Similarly, traditional load-balancing is based on local link utilization information, which may cause some paths overloaded while others remain underutilized. The local view of network status prevents precise and globally optimized decisions and adjustments. It would be helpful to send fast network notifications to upstream nodes which can perform action based on the view of regional or global network conditions. Overhead and Scalability Challenges: The distribution of high-volume network operational status information or frequent signaling introduces bandwidth and processing overhead. At scale, this becomes a bottleneck rather than a solution. IOAM and similar tools provide detailed telemetry information, but the collection and feedback loops are controller-centric. They cannot be used to deliver lightweight, real-time alerts for immediate action on specific network nodes. Carrying dynamic network state information in control protocols (e.g. routing protocols) also increases the overhead and churn of the control plane, which may have negative impact to the core functionality of the protocol. It would be useful to have solutions designed to avoid the overhead and churn introduced by telemetry flooding or route distribution, so it can adapt to large-scale networks and dynamic traffic patterns (e.g. AI workloads, cloud WAN bursts).

In particular, failure-detection and state-change propagation must occur within a timeframe short enough to meaningfully influence traffic engineering and load-balancing decisions before congestion or micro-loops develop. In backbone or datacenter networks, this typically implies a target of approximately 1-5 ms for notification delivery, with some environments requiring sub-millisecond performance. The precise requirement is driven by: the speed at which traffic shifts can induce overload the granularity of TE tuning (fine-grained vs. coarse-grained) the propagation diameter of the network the responsiveness of the control-plane and forwarding-plane components Therefore, this draft focuses on mechanisms capable of operating within these millisecond/sub-millisecond ranges, rather than mechanisms whose latency spans tens or hundreds of milliseconds, which are insufficient for preventing transient overload under rapid traffic transitions.

Consider a large-scale AI/ML training job distributed across multiple data centers. These clusters exchange terabits per second of data between GPU nodes, requiring ultra-low latency and high throughput to maintain synchronization.

As depicted in the above figure, a single fiber link failure or severe congestion event can disrupt the entire training run, leading to: Delays in job completion (hours to days for large models) Massive energy and compute cost waste due to resynchronization Degraded convergence accuracy if synchronization windows are missed

Today's mechanisms provide partial solutions but are not fast or precise enough for these scenarios: BFD : Provides fast forwarding path failure detection. It can be used for both link and path failure detection, while it cannot be used to detect link or path congestion, nor can it notify the failure or congestion to other nodes in the network. BFD is preconfigured with periodic message exchange, while fast notifications needs to be event-driven. When the transmit interval is set to a small value (e.g., at the level of ms), frequent BFD message exchange may become a burden to some systems. FRR /Route convergence: Without fast notification, the failure detection can take tens of milliseconds, followed by either local repair (FRR) or route convergence. The former lacks of global network situation thus may cause congestion on the backup paths, while the latter may breach strict synchronization deadlines. ECN : Provides binary congestion feedback to the endpoints, which is insufficient for granular congestion spikes on high-speed links, and the action can be slow. Telemetry (e.g., IOAM): Offers detailed information, but relies on collection and RTT-based feedback, which delays action. Flow control at the receiver/sender: Tied to RTT or packet loss, unsuitable for the bursty nature of AI traffic patterns. In practice, this means that by the time a fiber link failure is detected and recovery mechanisms are invoked, critical GPU synchronization barriers may already be missed, forcing rollbacks or restarts of the training process.

Fast network notification mechanisms could improve the response to fiber link failures and congestion in distributed AI/ML clusters: Real-Time Alerts: Nodes adjacent to the failure or congestion could immediately (e.g., at 10 ms level) send lightweight notifications to nodes whose fowarding paths can be affected. Action-Oriented Response: Upon receiving the notification, routing and load balancing mechanisms could instantly shift traffic to backup paths or alternative DC interconnects. Granularity: Notifications could carry more detailed information than "link failure/congestion," e.g., indicating specific link utilization, queue buildup or microburst congestion, allowing differentiated responses to different traffic flows. Complementary: The fast notification solutions are complementary to BFD, FRR or Telemetry, it would bridge the time gap between event onset and slower control plane or telemetry-driven responses, and enable network-wide optimization. By deploying fast notifications, large AI/ML workloads can maintain synchronization across data centers even during transient failures or congestion, protecting job completion time and resource utilization. Existing Approach: BFD detects failure after tens of ms FRR may cause congestion on backup paths Reroute/convergence delays impact GPU sync Result: Training stalls, compute resources wasted, job completion delayed Fast Notifications Approach: BFD detects failure after tens of ms Fast notification alerts upstream nodes of failure or congestion in real time Regional or global TE steers traffic quickly to alternate link/path without causing new congestion Result: Training continues with minimal disruption

The information carried in the fast network notifications, by the originating node, can be one or multiple of the following: Event Type: This can be used to indicate the type of events (e.g. failure, congestion, performance degradation, etc.). Location of Event: This can be used to indicate the location where the event occurred in the network (e.g. the identifier of the link, the node, or the queue, etc.). Fine-grained Network Status information: This can include quantifiable network metrics like link utilization, queue length, level of congestion, link or node delay, jitter, packet loss, etc. Path Identification information: This can be used to indicate the path which is affected by the event. Flow Identification information: This can include the identification or the 5-tuple of a flow which is affected by the event. Other information related to the network status change and need to be timely actioned may also be carried in the fast network notifications. Thus there is a need to work on the information model of fast network notifications to better understand what needs to be carried in the notifications.

Fast network notifications may be consumed by two broad forms of recipients: (1) recipient nodes that participate directly in forwarding or signaling, and (2) functions and applications that consume notifications in order to optimize, monitor, or adapt behaviors as depicted in the following two tables. Separating these categories clarifies which entities are physical/logical nodes versus which are higher-level functional consumers:

As illustrated in the figure above, the latency sensitivity of recipients decreases as one moves from the data plane to the application plane. Recipient nodes (e.g., adjacent forwarding elements, ingress routers, etc.) often require near-instantaneous notification, while functions and applications (e.g., routing protocols, analytics, NMS, etc.) may tolerate slightly longer timescales but still benefit from rapid awareness compared to existing mechanisms. The range of recipients of the notification depends on the type of recipients, it also depends on what type of action is required. The mechanism to determine the type and range of the recipients is something needs further consideration.

Depending on the position and number of the recipient nodes, fast network notifications may be sent via one of the following delivery modes: Unicast directly to the recipient node Multicast to a group of recipient nodes Hop-by-hop to a series of receipt nodes along a specified path Flooding in a specified range of the network Additionally, recipient nodes or functions may subscribe to specific types of notifications based on their roles or interests. A subscription-based approach enables selective delivery, reduces unnecessary signaling overhead, and ensures that each recipient receives only the information relevant to its function. Mechanisms supporting both delivery and subscription must guarantee timely, reliable, and secure propagation of notifications. Examples: Adjacent routers subscribing to all local failure notifications Centralized controllers subscribing only to congestion alerts exceeding defined thresholds Applications or analytics systems subscribing to performance degradation events affecting specific flows or services The mechanisms to support the above delivery mode needs to make sure the notification is always sent to the targeted recipient noded in a timely manner. It could be based on existing messaging and transport mechanisms, or a new protocol may be introduced.

Once a fast network notification is received, the recipient needs to take appropriate actions to help mitigating the event reported in the fast network notification. The action can be based on the information carried in the fast network notification, or it can be based on both the information of the notification and the information obtained by the recipient in other ways. The action to be performed by the recipient may be explicitly carried in the notification, or it may be implicitly determined by the type of information carried in the notification. Some actions are mandatory, while some actions can be optional. The possible actions to the notification can be but not limit to one or multiple of the following: Switch all traffic from a path to other available paths Steers specific traffic flows to alternate links or paths Modifies the load balancing ratio among a group of paths Sends the notification further to other recipients Whether actions need to be explicitly indicated in the notification, and if so, which ones, requires further consideration.

Current network mechanisms were not designed for the responsiveness and scale required by todays' dynamic environments. Techniques such as load balancing, protection switching, and flow control rely on telemetry and feedback loops that are often too slow, too coarse, or too resource-intensive. This results in performance bottlenecks, delayed recovery, and inefficiencies in large-scale AI, cloud, and WAN deployments. A fast network notification mechanism could help to address these gaps by providing lightweight, real-time, actionable alerts that complement existing tools and enable faster, more accurate network management decisions.

This document has no IANA actions.

Fast network notifications, if not properly authenticated and rate-limited, could be exploited as a vector for Denial-of-Service (DoS) attacks. An attacker able to inject or flood spurious notifications may trigger unnecessary re-convergence, path changes or repeated state updates, overwhelming both recipient nodes and higher-level applications. Implementations must therefore ensure integrity protection, origin authentication, and appropriate rate controls on notification messages.

The authors would like to thank Alia Atlas, David Black, Jeffrey Haas, Tony Li and Carlos J. Bernardos for their valuable comments and discussion.

The following people contributed substantially to the content of this document.