From 8685f9aecd97a2d13573cd5541bdab3b6c957a4b Mon Sep 17 00:00:00 2001 From: Brian Barrett Date: Wed, 19 Sep 2018 22:10:28 +0000 Subject: [PATCH] openib: Disable CUDA async by default Disable async receive for CUDA under OpenIB. While a performance optimization, it also causes incorrect results for transfers larger than the GPUDirect RDMA limit. This change has been validated and approved by Akshay. References #3972 Signed-off-by: Brian Barrett (cherry picked from commit 9344afd485d23c41a18ab57d6df8550826c24b9e) Signed-off-by: Brian Barrett --- opal/mca/btl/openib/btl_openib_mca.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/opal/mca/btl/openib/btl_openib_mca.c b/opal/mca/btl/openib/btl_openib_mca.c index 5cffd240c21..2e5acf7e3cc 100644 --- a/opal/mca/btl/openib/btl_openib_mca.c +++ b/opal/mca/btl/openib/btl_openib_mca.c @@ -560,7 +560,7 @@ int btl_openib_register_mca_params(void) CHECK(reg_bool("cuda_async_recv", NULL, "Enable or disable CUDA async recv copies " "(true = async; false = sync)", - true, &mca_btl_openib_component.cuda_async_recv)); + false, &mca_btl_openib_component.cuda_async_recv)); /* Also make the max send size larger for better GPU buffer performance */ mca_btl_openib_module.super.btl_max_send_size = 128 * 1024; /* Turn of message coalescing - not sure if it works with GPU buffers */