From 447ab5cca784de1da00a94a613846c37bd53bfe8 Mon Sep 17 00:00:00 2001 From: Wenkai Du Date: Tue, 17 Dec 2024 09:26:51 -0800 Subject: [PATCH] Enable HDP and GDRCOPY flush by default --- src/transport/net.cc | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/src/transport/net.cc b/src/transport/net.cc index ca2ddce5c..882d32e58 100644 --- a/src/transport/net.cc +++ b/src/transport/net.cc @@ -230,7 +230,7 @@ static ncclResult_t sendSetup(struct ncclComm* comm, struct ncclTopoGraph* graph // GDRCOPY support: TAIL_ENABLE When enabled locates the RX proxy tail in CUDA memory NCCL_PARAM(GdrCopySyncEnable, "GDRCOPY_SYNC_ENABLE", 1); // GDRCOPY support: FLUSH_ENABLE When enabled uses a PCI-E read to flush GDRDMA buffers -NCCL_PARAM(GdrCopyFlushEnable, "GDRCOPY_FLUSH_ENABLE", 0); +NCCL_PARAM(GdrCopyFlushEnable, "GDRCOPY_FLUSH_ENABLE", 1); /* Setup recv connector */ static ncclResult_t recvSetup(struct ncclComm* comm, struct ncclTopoGraph* graph, struct ncclPeerInfo* myInfo, struct ncclPeerInfo* peerInfo, struct ncclConnect* connectInfo, struct ncclConnector* recv, int channelId, int connIndex) { @@ -1355,7 +1355,7 @@ static ncclResult_t sendProxyProgress(struct ncclProxyState* proxyState, struct } RCCL_PARAM(NetHdpFlush, "NET_HDP_FLUSH", 1); -RCCL_PARAM(NetGdrFlush, "NET_GDR_FLUSH", 1); +RCCL_PARAM(NetGdrFlush, "NET_GDR_FLUSH", 0); static ncclResult_t recvProxyProgress(struct ncclProxyState* proxyState, struct ncclProxyArgs* args) { #if defined(ENABLE_NPKIT) && defined(ENABLE_NPKIT_NET_COLLECT_POLL_CNT)