Reproduce “Avoiding Cross-Datacenter Collective Congestion via Disaggregated Buffering” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mariano Scazzariello, Noga H. Rotman, Dima Gavrilenko, Sajy Khashab, Alexander Shpiner, Matty Kadosh, Marco Chiesa, Dejan Kostic and 1 more

SPILLWAY, a transparent in-network mechanism, absorbs cross-datacenter traffic collisions by deflecting packets to disaggregated buffers at the destination DC boundary, reducing microbatch completion time by up to 14.7% and overall iteration time by approximately 5.3% in large language model training.

Key insights

  • Disaggregated buffering positioned at the destination datacenter boundary enables microsecond-scale control over cross-DC congestion, a granularity unattainable by millisecond-latency cross-DC congestion controllers.
  • Leveraging 'deflect-on-drop' as an implicit feedback signal allows spillway nodes to intelligently re-inject buffered packets without direct network visibility, by waiting for a 'quiet interval' and probing the network.
  • Combining anycast for initial load balancing of deflected traffic and a 'sticky' unicast strategy for subsequent retransmissions to the same spillway is crucial for efficient resource utilization and maintaining flow context.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.