rdma: Support early completion of recv() requests

include/nccl_ofi_param.h

@@ -396,6 +396,14 @@ OFI_NCCL_PARAM_INT(use_low_lat_tc, "USE_LOW_LATENCY_TC", 1);
  */
 OFI_NCCL_PARAM_INT(force_num_rails, "FORCE_NUM_RAILS", 0);
 
+/*
+ * 1 to enable early completion, 0 to disable it.
+ * Default at -1 to follow the data progress model, given that 
+ * early completion feature is contigent on FI_PROGRESS_AUTO data progress model
+ * i.e. enabled when FI_PROGRESS_AUTO, otherwise disabled
+ */
+OFI_NCCL_PARAM_INT(early_completion, "EARLY_COMPLETION", -1);
+
 #ifdef __cplusplus
 } // End extern "C"
 #endif

include/nccl_ofi_rdma.h

@@ -104,6 +104,7 @@ enum nccl_ofi_rdma_msg_type {
 	NCCL_OFI_RDMA_MSG_CTRL,
 	NCCL_OFI_RDMA_MSG_EAGER,
 	NCCL_OFI_RDMA_MSG_CLOSE,
+	NCCL_OFI_RDMA_MSG_CTRL_NO_COMPLETION,
 	NCCL_OFI_RDMA_MSG_INVALID = 15,
 	NCCL_OFI_RDMA_MSG_MAX = NCCL_OFI_RDMA_MSG_INVALID,
 };
@@ -264,6 +265,8 @@ typedef struct {
 	/* Total number of completions. Expect one completion for receiving the
 	 * control message and one completion for each send segment. */
 	int total_num_compls;
+	/* True to use fi_write instead of fi_writedata in send() */
+	bool use_fi_write;
 #if HAVE_NVTX_TRACING
 	nvtxRangeId_t trace_id;
 	nvtxRangeId_t seg_trace_id[MAX_NUM_RAILS];

src/nccl_ofi_api.c

@@ -789,14 +789,6 @@ ncclResult_t nccl_net_ofi_irecv_v9(void* recvComm, int n, void** data,
 		return check_return(ncclInvalidArgument);
 	}
 
-	/* 
-	 * Reset to NULL for now until optional receive completion logic is
-	 * implemented
-	 */
-	if (*request == (void *)NCCL_NET_OPTIONAL_RECV_COMPLETION) {
-		*request = NULL;
-	}
-
 	ncclResult_t validation_result = msg_length_verify_max_size(sizes, n);
 	if (validation_result != ncclSuccess) {
 		return check_return(validation_result);

src/nccl_ofi_rdma.c

@@ -107,6 +107,8 @@ static bool is_max_write_inline_size_initialized = false;
 /* CPU cache line size */
 static ssize_t cpu_cache_line_size;
 
+static bool early_completion = false;
+
 /* Function prototypes */
 static int send_progress(nccl_net_ofi_rdma_req_t *req);
 
@@ -978,6 +980,7 @@ static inline int update_send_data_from_remote(nccl_net_ofi_rdma_send_comm_t *s_
 	send_data->wdata =
 		GET_RDMA_WRITE_IMM_DATA(s_comm->remote_comm_id, req->msg_seq_num, send_data->schedule->num_xfer_infos);
 
+	send_data->use_fi_write = (ctrl_msg->type == NCCL_OFI_RDMA_MSG_CTRL_NO_COMPLETION);
 	return 0;
 }
 
@@ -1378,6 +1381,8 @@ static inline int handle_rx_buff_recv(nccl_net_ofi_rdma_device_t *device, int ra
 			goto exit;
 		}
 		break;
+	case NCCL_OFI_RDMA_MSG_CTRL_NO_COMPLETION:
+		/* fall through to NCCL_OFI_RDMA_MSG_CTRL case */
 	case NCCL_OFI_RDMA_MSG_CTRL:
 		/* CTRL receive completion */
 		assert(cq_entry->len == nccl_net_ofi_rdma_ctrl_msg_size(ep->num_rails, ep->use_long_rkeys));
@@ -3344,7 +3349,8 @@ static inline int insert_send_ctrl_req(
 				int dev_id, uint16_t msg_seq_num, void *buff,
 				size_t size,
 				nccl_net_ofi_rdma_mr_handle_t *buff_mr_handle,
-				nccl_net_ofi_rdma_req_t *recv_req)
+				nccl_net_ofi_rdma_req_t *recv_req,
+				bool recv_completion_optional)
 {
 	nccl_net_ofi_scheduler_t *scheduler = device->scheduler;
 	nccl_net_ofi_rdma_ep_t *ep = (nccl_net_ofi_rdma_ep_t *)r_comm->base.base.ep;
@@ -3406,7 +3412,8 @@ static inline int insert_send_ctrl_req(
 
 	nccl_net_ofi_rdma_ctrl_msg_t *ctrl_msg = rdma_send_ctrl_get_msg(send_ctrl_data);
 
-	ctrl_msg->type = NCCL_OFI_RDMA_MSG_CTRL;
+	/* If early completion is turned on, CTRL msg type will be NCCL_OFI_RDMA_MSG_CTRL_NO_COMPLETION to influence send() behavior */
+	ctrl_msg->type = recv_completion_optional ? NCCL_OFI_RDMA_MSG_CTRL_NO_COMPLETION : NCCL_OFI_RDMA_MSG_CTRL;
 	ctrl_msg->remote_comm_id = r_comm->remote_comm_id;
 	ctrl_msg->msg_seq_num = msg_seq_num;
 	ctrl_msg->buff_addr = (uint64_t)buff;
@@ -3482,7 +3489,8 @@ static inline int allocate_rdma_recv_req(
 				int dev_id, uint16_t msg_seq_num, void *buff,
 				size_t size,
 				nccl_net_ofi_rdma_mr_handle_t *buff_mr_handle,
-				nccl_net_ofi_rdma_req_t **ret_req)
+				nccl_net_ofi_rdma_req_t **ret_req,
+				bool recv_completion_optional)
 {
 	int ret = 0;
 	rdma_req_recv_data_t *recv_data;
@@ -3503,14 +3511,15 @@ static inline int allocate_rdma_recv_req(
 	req->msg_seq_num = msg_seq_num;
 
 	recv_data = get_recv_data(req);
-	recv_data->total_num_compls = 2;
+	/* In the case of early completion, only expect the completion for control msg itself */
+	recv_data->total_num_compls = recv_completion_optional ? 1 : 2;
 	recv_data->eager_copy_req = NULL;
 	recv_data->dst_buff = buff;
 	recv_data->dst_len = size;
 	recv_data->dest_mr_handle = buff_mr_handle;
 
 	/* TODO consolidate arguments to insert_send_ctrl_req and insert_recv_segms_req */
-	ret = insert_send_ctrl_req(r_comm, device, dev_id, msg_seq_num, buff, size, buff_mr_handle, req);
+	ret = insert_send_ctrl_req(r_comm, device, dev_id, msg_seq_num, buff, size, buff_mr_handle, req, recv_completion_optional);
 	if (ret) {
 		NCCL_OFI_WARN("Failed to insert send ctrl request into recv request");
 		return ret;
@@ -3606,9 +3615,14 @@ static int recv(nccl_net_ofi_recv_comm_t *recv_comm, int n, void **buffers,
 	nccl_net_ofi_rdma_mr_handle_t **mr_handles = (nccl_net_ofi_rdma_mr_handle_t **)mhandles;
 	uint16_t msg_seq_num = 0;
 	bool eager = false;
+	bool recv_completion_optional = false;
 
 	assert(r_comm != NULL);
 
+	if (early_completion && *base_req == (void *)NCCL_NET_OPTIONAL_RECV_COMPLETION) {
+		recv_completion_optional = true;
+	}
+
 	if (r_comm->comm_active == false) {
 		NCCL_OFI_WARN("Called irecv on inactive communicator");
 		ret = -EINVAL;
@@ -3677,7 +3691,7 @@ static int recv(nccl_net_ofi_recv_comm_t *recv_comm, int n, void **buffers,
 
 	ret = allocate_rdma_recv_req(r_comm, device, dev_id, msg_seq_num,
 					buffers[0], sizes[0],
-					mr_handles[0], &req);
+					mr_handles[0], &req, recv_completion_optional);
 	if (ret != 0) {
 		goto error;
 	}
@@ -5499,7 +5513,8 @@ static int post_rma_write(nccl_net_ofi_rdma_req_t *req)
 
 static int post_rdma_write(nccl_net_ofi_rdma_req_t *req,
 			   nccl_net_ofi_rdma_send_comm_rail_t *comm_rail,
-			   nccl_net_ofi_xfer_info_t *xfer_info)
+			   nccl_net_ofi_xfer_info_t *xfer_info,
+			   bool use_fi_write)
 {
 	rdma_req_send_data_t *send_data = get_send_data(req);
 	assert(xfer_info->rail_id < send_data->buff_mr_handle->num_rails);
@@ -5509,12 +5524,19 @@ static int post_rdma_write(nccl_net_ofi_rdma_req_t *req,
 
 	ssize_t rc;
 	/* Post RDMA write */
-	rc = fi_writedata(comm_rail->local_ep, (void*)((uintptr_t)send_data->buff + xfer_info->offset),
-				xfer_info->msg_size, desc, send_data->wdata,
-				comm_rail->remote_addr,
-				send_data->remote_buff + xfer_info->offset,
-			  send_data->remote_mr_key[rail_id], (void *)&req->ctx[rail_id]);
-
+	if (use_fi_write) {
+		rc = fi_write(comm_rail->local_ep, (void*)((uintptr_t)send_data->buff + xfer_info->offset),
+					xfer_info->msg_size, desc,
+					comm_rail->remote_addr,
+					send_data->remote_buff + xfer_info->offset,
+					send_data->remote_mr_key[rail_id], (void *)&req->ctx[rail_id]);
+	} else {
+		rc = fi_writedata(comm_rail->local_ep, (void*)((uintptr_t)send_data->buff + xfer_info->offset),
+					xfer_info->msg_size, desc, send_data->wdata,
+					comm_rail->remote_addr,
+					send_data->remote_buff + xfer_info->offset,
+					send_data->remote_mr_key[rail_id], (void *)&req->ctx[rail_id]);
+	}
 	if ((rc != 0) && (rc != -FI_EAGAIN)) {
 		NCCL_OFI_WARN("fi_writedata failed; RC: %zd, Error: %s",
 			      rc, fi_strerror(-rc));
@@ -5643,7 +5665,7 @@ static int send_progress(nccl_net_ofi_rdma_req_t *req)
 				nccl_net_ofi_rdma_send_comm_rail_t *comm_rail =
 					rdma_send_comm_get_rail(s_comm, xfer_info->rail_id);
 
-				ret = post_rdma_write(req, comm_rail, xfer_info);
+				ret = post_rdma_write(req, comm_rail, xfer_info, send_data->use_fi_write);
 
 				if (ret == 0) // Successfully sent the xfer with this rail
 					send_data->xferred_rail_id++;
@@ -8115,6 +8137,44 @@ int nccl_net_ofi_rdma_init(const char *provider_filter,
 		goto error;
 	}
 
+	/* 
+	* NCCL Net v9 API Optimization for LL/LL128 Protocols
+	* 
+	* Background:
+	* When using LL (Low Latency) or LL128 protocols, NCCL sets the request pointer 
+	* to NCCL_NET_OPTIONAL_RECV_COMPLETION in irecv() calls. This indicates that 
+	* the plugin can complete a receiver request early without plugin explicitly
+	* polling the CQ to validate data arrival. This is achievable because NCCL itself
+	* following LL protocol semantics will validate data arrival by checking the flag bytes.
+	*
+	* Plugin Optimization Details:
+	* 1. Receiver Side:
+	*    - Marks request completion immediately after CTRL message send completion
+	*    - Does not wait for RDMA write operation completion
+	*
+	* 2. Sender Side:
+	*    - Uses fi_write instead of fi_writedata, to eliminate unnecessary CQ entries on RX side
+	*
+	* Requirements:
+	* - Provider must use FI_PROGRESS_AUTO data progress model
+	*/
+	switch (ofi_nccl_early_completion()) {
+		case -1: 
+			early_completion = data_progress_auto;
+			break;
+		case 1:
+			if (!data_progress_auto) {
+				NCCL_OFI_WARN("Early completion enablement failed due to provider data progress model is not FI_PROGRESS_AUTO");
+				ret = -ENOTSUP;
+				goto error;
+			}
+			early_completion = true;
+			break;
+		case 0:
+			early_completion = false;
+			break;
+	}
+
 	/* Create NCCL OFI topology */
 	topo = nccl_ofi_topo_create(provider_list);
 	if (!topo) {