Avoiding Cross-Datacenter Collective Congestion via Disaggregated Buffering

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Scazzariello, Mariano, Rotman, Noga H., Gavrilenko, Dima, Khashab, Sajy, Shpiner, Alexander, Kadosh, Matty, Chiesa, Marco, Kostic, Dejan, Silberstein, Mark
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917490346426368
author Scazzariello, Mariano
Rotman, Noga H.
Gavrilenko, Dima
Khashab, Sajy
Shpiner, Alexander
Kadosh, Matty
Chiesa, Marco
Kostic, Dejan
Silberstein, Mark
author_facet Scazzariello, Mariano
Rotman, Noga H.
Gavrilenko, Dima
Khashab, Sajy
Shpiner, Alexander
Kadosh, Matty
Chiesa, Marco
Kostic, Dejan
Silberstein, Mark
contents LLM training at the scale of tens of thousands of GPUs now spans multiple datacenters (DC), making cross-DC collectives over long-haul links unavoidable. A critical and overlooked bottleneck arises when these collectives collide with intra-DC traffic at the destination - a common pattern in real workloads. The multi-millisecond congestion control loop is too slow to react, triggering severe packet loss and congestion collapse. We present Spillway, a transparent in-network mechanism that buffers dropped packets in switch-disaggregated buffers in a destination data center and drains them once congestion subsides. Through large-scale end-to-end simulations and a hardware prototype, we show that Spillway eliminates performance degradation from collective collisions, reducing iteration time by up to 14 %, without changes to end hosts or training frameworks.
format Preprint
id arxiv_https___arxiv_org_abs_2605_11852
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Avoiding Cross-Datacenter Collective Congestion via Disaggregated Buffering
Scazzariello, Mariano
Rotman, Noga H.
Gavrilenko, Dima
Khashab, Sajy
Shpiner, Alexander
Kadosh, Matty
Chiesa, Marco
Kostic, Dejan
Silberstein, Mark
Networking and Internet Architecture
LLM training at the scale of tens of thousands of GPUs now spans multiple datacenters (DC), making cross-DC collectives over long-haul links unavoidable. A critical and overlooked bottleneck arises when these collectives collide with intra-DC traffic at the destination - a common pattern in real workloads. The multi-millisecond congestion control loop is too slow to react, triggering severe packet loss and congestion collapse. We present Spillway, a transparent in-network mechanism that buffers dropped packets in switch-disaggregated buffers in a destination data center and drains them once congestion subsides. Through large-scale end-to-end simulations and a hardware prototype, we show that Spillway eliminates performance degradation from collective collisions, reducing iteration time by up to 14 %, without changes to end hosts or training frameworks.
title Avoiding Cross-Datacenter Collective Congestion via Disaggregated Buffering
topic Networking and Internet Architecture
url https://arxiv.org/abs/2605.11852