Dirk Kutscher

Today’s transport protocols schedule packets without understanding computation. SAFS introduces receiver-driven semantic-flow scheduling, allowing distributed AI runtimes to expose computation semantics so that computation-critical communication is prioritized before pipeline bubbles can form.

Modern AI systems increasingly execute as pipelines spanning multiple GPUs, servers, and disaggregated components. In these systems, network communication is no longer just a data transfer problem – it directly determines whether expensive GPU resources remain busy or become idle.

Yet today's datacenter transports remain largely application-agnostic. They optimize congestion, flow completion time, packet priorities, or deadlines, but they do not understand why a particular communication flow matters to the computation.

The figure below illustrates the core idea behind SAFS. Conventional transport schedules communication independently of application progress, whereas SAFS uses lightweight runtime semantics to prioritize computation-critical communication before the no-bubble deadline.

Our accepted IEEE ICNP 2026 paper introduces SAFS (Slack-Aware Flow Scheduling), a transport architecture that schedules communication according to AI computation semantics.

The Key Observation

A network transfer is not harmful simply because it is delayed.

Instead, it becomes harmful only when it arrives after the downstream computation has exhausted the slack that could otherwise have hidden the communication. At that point, the delay becomes an exposed pipeline bubble, leaving GPUs idle while waiting for missing data.

This leads to a fundamentally different transport objective.

Instead of asking:

Which flow should finish first?

SAFS asks:

Which communication dependency will become the next computation bottleneck?

From Packets to Semantic Flows

SAFS introduces the notion of semantic flows.

Rather than treating network traffic as anonymous byte streams, each communication flow is associated with lightweight runtime metadata describing its role within the computation. This includes information such as:

• remaining data volume,
• a no-bubble timestamp indicating when the data must arrive,
• dependency information,
• and runtime updates as execution progresses.

Receivers use this semantic information to allocate bandwidth to the communication flows that are most critical for maintaining continuous pipeline execution.

Receiver-Driven Scheduling for AI Workloads

SAFS combines these computation semantics with a receiver-driven transport architecture.

Instead of allocating bandwidth solely according to congestion, flow size, or externally assigned deadlines, receivers compute the transmission rates required to keep each communication dependency on its no-bubble trajectory.

This allows application semantics to guide transport scheduling while existing transport mechanisms continue to provide congestion control, fairness, and reliable data delivery.

Because SAFS augments rather than replaces existing transport protocols, it can be deployed incrementally without requiring changes to AI applications or accelerator hardware.

Evaluation

We demonstrate SAFS in two representative AI environments:

• Pipeline-parallel training, where communication between pipeline stages competes with background traffic.
• Disaggregated LLM inference, where hidden-state transfers determine when feed-forward computation can begin.

Across both workloads, SAFS consistently prioritizes the communication dependencies that determine end-to-end execution performance.

Highlights include:

• Up to 38.7% lower iteration time during ZeroBubble pipeline training under background traffic.
• 38.7% lower p99 TPOT than D2TCP for large-scale disaggregated inference.
• 66.2% lower p99 TPOT than Homa.
• Zero urgent deadline misses in targeted receiver-incast microbenchmarks.

Why It Matters

SAFS explores a broader research direction:

Transport protocols should understand computation—not just packets.

This represents a fundamental shift from optimizing packet delivery toward optimizing application progress. As AI systems become increasingly distributed and disaggregated, exposing computation semantics to the transport layer enables the network to prioritize the communication that actually determines application performance.

Rather than optimizing generic network metrics, SAFS directly optimizes what distributed AI workloads ultimately care about: continuous computation, high accelerator utilization, and fewer pipeline bubbles.

Reference

Ruilong Wu, Yisu Wang, Jiaxiang Zou, Xiangsen Chen, Yixiang Zhu, Xinjiao Li, Xinyu Chen, Dirk Kutscher; SAFS: Slack-Aware Flow Scheduling for Pipelined AI Workloads; to appear at the 34th IEEE International Conference on Network Protocols (ICNP); Tempe, Arizona, USA, October 05-08, 2026

Organizers

• Antoine Fressancourt
• Dirk Kutscher

The 1st workshop on Inter-networking challenges for AI (INet4AI), collocated with ACM CoNEXt'25, was held on the 1st of December 2025 in Hong-Kong. The workshop was inspired by ongoing discussion in the IRTF on research challenges for (inter-)networking technologies for AI workloads.

This full day workshop explored some of the networking challenges of large-scale distributed AI workloads in environments, characterized by node and network heterogeneity, as well as dynamically changing resource availability and utilization. During this inaugural edition, researchers from both academia (HKUST, ETH Zurich, Politecnico di Milano, University of Napoli, Tsinghua University, TU Munich) and industry (Huawei, AMD, Microsoft, and others) discussed possible solutions to address the challenges raised by Internet-scale distributed AI systems with four workshop paper presentations and three invited talks. In this report, we will first give a summary of the workshop papers and invited talks. Then we will draw some general remarks regarding the ongoing efforts done in our community to address INet4AI challenges.

Check out the full report.

Program Overview

• Invited talk — Tommaso Bonato — Uno: A One-Stop Solution for Inter- and Intra-Datacenter Congestion Control and Reliable Connectivity — Paper · Slides.
• AI4Net paper — Shaked Leibzirer — Self-supervised Application-level Network Traffic Inversion — Paper · Slides.
• Net4AI paper — German Sviridov — Latency-Optimal Load Balancing For Distributed MoE Inference — Paper.
• Invited talk — Mingxing Zhang — From Homogeneous to Disaggregated Architectures for Large Model Inference — Slides.
• Net4AI paper — Jiaheng Xiong — SCALE-CCL: A Scalable Collective Communication Library for Wide-Area Distributed Training — Paper · Slides.
• Net4AI paper — Giuseppe Aceto — You’ve got a few GPUs, now what?! — Experimenting with a Nano-Cluster for Distributed Training of AI Models — Paper · Slides.
• Invited talk — Wenjia Wei — Debriefing the Open Innovation Platform for UnifiedBus — Slides.