pytorch
f993e5ac - [1.6] Update TensorPipe submodule (#40634)

Commit
4 years ago
[1.6] Update TensorPipe submodule (#40634) Upstream PR: #40614 Summary: This update pulls in a oneliner fix, which sets the TCP_NODELAY option on the TCP sockets of the UV transport. This leads to exceptional performance gains in terms of latency, with about a 25x improvement in one simple benchmark. This thus resolves a regression that TensorPipe had compared to the ProcessGroup agent and, in fact, ends up beating it by 2x. The benchmark I ran is this, with the two endpoints pinned to different cores of the same machine: ``` torch.jit.script def remote_fn(t: int): return t torch.jit.script def local_fn(): for _ in range(1_000_000): fut = rpc.rpc_async("rhs", remote_fn, (42,)) fut.wait() ``` And the average round-trip time (one iteration) is: - TensorPipe with SHM: 97.2 us - TensorPipe with UV _after the fix_: 205us - Gloo: 440us - TensorPipe with UV _before the fix_: 5ms Test Plan: Ran PyTorch RPC test suite
Author
lw lw
Parents
Loading