[Feature] Cross-node communication without InfiniBand

Question

[Feature] Cross-node communication without InfiniBand

chenhongyu2048 opened this issue a month ago · 4 comments

chenhongyu2048 commented a month ago

I wonder if msccclpp supports communication between GPUs on two nodes that don't have IB, e.g. in the case of ethernet-only.

Answer 1 · 2024-12-12T02:27:54.000Z

We do support both RoCE and TCP.

Answer 2 · 2024-12-12T02:42:37.000Z

Thanks, but I'm still wondering how I should use TCP/Ethernet, should I set Transport to Ethernet to do so?
And in #408 you answer that SmChannel can only use cudaipc, so if I want to use Ethernet, should I choose ProxyChannel?

Answer 3 · 2024-12-15T03:56:29.000Z

Yes, ProxyChannel with Ethernet transport will run over TCP.

Answer 4 · 2024-12-15T08:15:54.000Z

Thanks for your reply!