Production RDMA needs software-layer management, not just hardware offload

ByteDance's ByteExpress layer cuts parameter fetch latency by 43.7% and push latency by 74% versus kernel TCP in production ML workloads.

Chinese Tech
Yibo Huang · Yi Cui · Yukai Huang · Huaping Zhou · Jiancong Tong · Zhaoyu Wu · +9 more

Fudan University · ByteDance Inc.

Research Digest··3 min read
The authors report on ByteExpress, a software communication layer over hardware RDMA deployed in ByteDance's parameter-server inference and training systems.

ByteExpress sits between Archon, ByteDance's existing RPC framework, and RDMA NICs (RNICs) using Reliable Connection transport.

Why this paper

From ByteDance Inc. and Fudan University

In one line

ByteExpress, a software layer over RDMA, cuts inference fetch latency by 35.7% and training push latency by 74% versus TCP.

What we could check

  • ·No code link found
  • ·No weights link found
  • ·No dataset link found
  • ·No compute details found
  • ·No stated limitations found
  • ✓Reports numbers on named benchmarks (2 benchmarks)

Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.

§

Research Digest

Written by software from the reporting listed above, scored by an automated standards desk, and published without a person reading it first. If something here is wrong, tell the editor and it will be put right.

How we workSubscribe