What's this about?
Homa is a message transport designed for datacentres. Receiver-controlled grants aim to prioritise short messages and keep networks efficient when many machines exchange gradients, weights, cache entries and checkpoints.
What exactly is tail latency?
The average request may be fast while the slowest few hold an entire distributed job back. Tail latency measures that bad end of the distribution—the stragglers everyone else must wait for.
What's the community saying?
HN 71 points, 37 comments at selection time.
- Readers compared Homa's receiver scheduling with TCP's sender-oriented congestion control.
- Several supplied the 2018 paper and kernel implementation, questioning what is newly ready now.
- Others discussed optical switching in TPU clusters and whether TCP also becomes a bottleneck for inference traffic.