Swing: Short-cutting Rings for Higher Bandwidth Allreduce
Talkat 21st USENIX Symposium on Networked Systems Design and Implementation (NSDI '24) · 2024-04-16..04-18·
summary
New algorithm Swing shortens allreduce paths by short-cutting rings and swinging between torus directions, splitting each message so every node uses all local ports and alternates send directions to cut flow collisions. Evaluation uses SST Bucket Level simulation with 400 Gbit/s links and ~100 ns link latency, scales to 16k nodes, and shows up to 3x faster (≈+120%) for 32B–128MiB vectors, while the largest regression is about −22% for very large reductions.
AI-written summary · about
chapters (6)
- 0:00Introduction to Swing
- 1:10Torus networks in deep learning
- 2:45State of the art and challenges
- 7:16Algorithm intuition
- 10:10Efficient port utilization
- 12:26Evaluation and conclusion
related talks
16mOptiReduce: Resilient and Tail-Optimal AllReduce for Distributed Deep Learning in the...
45mFaster or better designed? Choose any two!
38mLost in Transduction
53mMapReduce Algorithmics
16mMinFlow: High-performance and Cost-efficient Data Passing for I/O-intensive Stateful...
54mReduction theorems for effective parameterized verification of round- based distributed algorithms
16mSöze: One Network Telemetry Is All You Need for Per-flow Weighted Bandwidth Allocation...
25mApproximating Fair Queueing on Reconfigurable Switches
