yezhengmao/nccl
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix GPU Direct RDMA detection.

Browse Source 
Whether the network supported GPU Direct RDMA or not was ignored,
causing sockets to break when cards were local enough that NCCL
tried to use it.
This commit is contained in:
Sylvain Jeaugey 2018-12-04 14:42:28 -08:00
parent b8a9a32ccb
commit 4b39a4cf91
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/transport/net.cu
View File

@ -241,7 +241,7 @@ static ncclResult_t netGetGdrSupport(int dev, int distance, int read, int* useGd
// Finally, check if the NIC supports it // Finally, check if the NIC supports it
int flags; int flags;
NCCLCHECK(ncclNetPtrSupport(dev, &flags)); NCCLCHECK(ncclNetPtrSupport(dev, &flags));
if (flags & NCCL_PTR_CUDA == 0) return ncclSuccess; if ((flags & NCCL_PTR_CUDA) == 0) return ncclSuccess;
*useGdr = 1; *useGdr = 1;
INFO(NCCL_INIT|NCCL_NET,"NET/%s : GPU Direct RDMA Enabled for GPU %d / HCA %d (distance %d >= %d), read %d", ncclNetName(), cudaDev, dev, distance, netGdrLevel, read); INFO(NCCL_INIT|NCCL_NET,"NET/%s : GPU Direct RDMA Enabled for GPU %d / HCA %d (distance %d >= %d), read %d", ncclNetName(), cudaDev, dev, distance, netGdrLevel, read);
return ncclSuccess; return ncclSuccess;