rdma/sendrecv: Lower thread requirement to DOMAIN

Lower the required threading limits to FI_THREAD_DOMAIN, with control progress set to FI_PROGRESS_CONTROL_UNIFIED. On Libfabric providers that use the utility completion queues, this will result in no locks for both the send/recv and cq polling calls inside Libfabric, since we already have domain-level exclusivity in the transports. Signed-off-by: Brian Barrett <[email protected]>
aws · Jan 28, 2025 · 0e0782e · 0e0782e
1 parent 94bc377
commit 0e0782e
Show file tree

Hide file tree

Showing 3 changed files with 31 additions and 11 deletions.
diff --git a/m4/check_pkg_libfabric.m4 b/m4/check_pkg_libfabric.m4
@@ -58,6 +58,7 @@ AC_DEFUN([CHECK_PKG_LIBFABRIC], [
                   FI_OPT_MAX_MSG_SIZE,
                   FI_OPT_SHARED_MEMORY_PERMITTED,
                   FI_MR_DMABUF,
+		  FI_PROGRESS_CONTROL_UNIFIED,
 		  FI_OPT_INJECT_RMA_SIZE],
                   [], [], [AC_INCLUDES_DEFAULT
 [#include <rdma/fi_endpoint.h>

diff --git a/src/nccl_ofi_rdma.c b/src/nccl_ofi_rdma.c
@@ -7728,14 +7728,22 @@ static void get_hints(struct fi_info *hints)
 	hints->domain_attr->mr_mode = FI_MR_LOCAL | FI_MR_HMEM | FI_MR_VIRT_ADDR |
 		FI_MR_ALLOCATED | FI_MR_PROV_KEY;
 	hints->domain_attr->mr_key_size = (size_t) ofi_nccl_mr_key_size();
-	hints->domain_attr->threading = FI_THREAD_SAFE;
-
-	/* Set progress mode to unspec to use the provider's default
-	 * mode.  We hard poll for completion, but if a provider is
-	 * faster with async progress, then we don't really care and
-	 * should let it do that. */
+	hints->domain_attr->threading = FI_THREAD_DOMAIN;
+
+	/* If libfabric is new enough to support
+	 * FI_PROGRESS_CONTROL_UNIFIED, specify MANUAL /
+	 * CONTROL_UNIFIED progress, to remove the domain lock from
+	 * the completion queue polling.  Otherwise, set
+	 * PROGRESS_UNSPEC to allow the provider to pick what it
+	 * thinks will go fastsest.
+	 */
+#if HAVE_DECL_FI_PROGRESS_CONTROL_UNIFIED
+	hints->domain_attr->control_progress = FI_PROGRESS_CONTROL_UNIFIED;
+	hints->domain_attr->data_progress = FI_PROGRESS_MANUAL;
+#else
 	hints->domain_attr->control_progress = FI_PROGRESS_UNSPEC;
-	hints->domain_attr->data_progress = FI_PROGRESS_UNSPEC;
+	hints->domain_attr->data_progress = FI_PROGRESS_UNSPEC;;
+#endif
 }
 
 

diff --git a/src/nccl_ofi_sendrecv.c b/src/nccl_ofi_sendrecv.c
@@ -2571,11 +2571,22 @@ static void sendrecv_get_hints(struct fi_info *hints, int req_gdr)
 
 	hints->ep_attr->type = FI_EP_RDM;
 
-	hints->domain_attr->threading = FI_THREAD_SAFE;
-
-	/* Set progress mode to unspec to use the provider's default mode. */
+	hints->domain_attr->threading = FI_THREAD_DOMAIN;
+
+	/* If libfabric is new enough to support
+	 * FI_PROGRESS_CONTROL_UNIFIED, specify MANUAL /
+	 * CONTROL_UNIFIED progress, to remove the domain lock from
+	 * the completion queue polling.  Otherwise, set
+	 * PROGRESS_UNSPEC to allow the provider to pick what it
+	 * thinks will go fastsest.
+	 */
+#if HAVE_DECL_FI_PROGRESS_CONTROL_UNIFIED
+	hints->domain_attr->control_progress = FI_PROGRESS_CONTROL_UNIFIED;
+	hints->domain_attr->data_progress = FI_PROGRESS_MANUAL;
+#else
 	hints->domain_attr->control_progress = FI_PROGRESS_UNSPEC;
-	hints->domain_attr->data_progress = FI_PROGRESS_UNSPEC;
+	hints->domain_attr->data_progress = FI_PROGRESS_UNSPEC;;
+#endif
 
 	/* Set MR mode bits to indicate FI_MR_BASIC registration */
 	hints->domain_attr->mr_mode |= FI_MR_VIRT_ADDR | FI_MR_ALLOCATED | FI_MR_PROV_KEY;