Skip to content

Commit 92689ce

Browse files
authored
Add NCCL_PROTO=^LL128 to recommended NCCL environment variables (#253)
1 parent 720fd50 commit 92689ce

File tree

1 file changed

+3
-0
lines changed

1 file changed

+3
-0
lines changed

docs/software/communication/nccl_env_vars

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -10,6 +10,9 @@ export NCCL_NET="AWS Libfabric"
1010
# https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html#nccl-net-gdr-level-formerly-nccl-ib-gdr-level
1111
export NCCL_NET_GDR_LEVEL=PHB
1212
export NCCL_CROSS_NIC=1
13+
# Starting with nccl 2.27 a new protocol (LL128) was enabled by default, which
14+
# typically performs worse on Slingshot. The following disables that protocol.
15+
export NCCL_PROTO=^LL128
1316
# These `FI` (libfabric) environment variables have been found to give the best
1417
# performance on the Alps network across a wide range of applications. Specific
1518
# applications may perform better with other values.

0 commit comments

Comments
 (0)