param: autogenerate topology file by default for RDMA protocol

Instead of using a static topology file. An autogenerated topology file
will be used when the user does not explicitly set a topology file
(`NCCL_TOPO_FILE`) and uses the RDMA protocol. For platform-aws builds,
this applies to the P5 platform.

Signed-off-by: Eric Raut <eraut@amazon.com>

include/nccl_ofi_param.h

@@ -162,13 +162,6 @@ OFI_NCCL_PARAM_STR(protocol, "PROTOCOL", NULL);
 
 OFI_NCCL_PARAM_INT(domain_per_thread, "DOMAIN_PER_THREAD", -1);
 
-/*
- * When enabled and RDMA communication protocol is used, write NCCL
- * topology file and set environment variable `NCCL_TOPO_FILE`. OFI plugin
- * writes the NCCL topology file to a memfd file.
- */
-OFI_NCCL_PARAM_INT(topo_file_write_enable, "TOPO_FILE_WRITE_ENABLE", 0);
-
 /*
  * Disable the native RDMA write support check when using the "RDMA" protocol
  * for send/recv operations on AWS platforms. When the check is disabled, the

src/nccl_ofi_rdma.c

@@ -247,10 +247,8 @@ static inline struct fid_domain *get_domain_from_endpoint(nccl_net_ofi_rdma_ep_t
 /*
  * @brief	Write topology to NCCL topology file
  *
- * If environment variable `OFI_NCCL_TOPO_FILE_WRITE_ENABLE` is set,
- * this function writes a NCCL topology file to a memfd file.
- *
- * It also sets environment variable `NCCL_TOPO_FILE` to the
+ * This function writes a NCCL topology file to a memfd file, and
+ * sets environment variable `NCCL_TOPO_FILE` to the
  * filename path of topology file.
  *
  * @param	topo
@@ -264,13 +262,6 @@ static int write_topo_file(nccl_ofi_topo_t *topo)
 	int topo_fd = -1;
 	FILE *file = NULL;
 
-	/* This function is a no-op in case writing topology file is not enabled explicitly */
-	if (!ofi_nccl_topo_file_write_enable()) {
-		NCCL_OFI_TRACE(NCCL_INIT | NCCL_NET,
-			       "Topology write not enabled; skipping");
-		goto exit;
-	}
-
 	/**
 	 * If `NCCL_TOPO_FILE` is already set, don't set it again.
 	 *

src/platform-aws.c

@@ -67,7 +67,7 @@ struct ec2_platform_data {
 	},
 	{
 		.name = "p5.48xlarge",
-		.topology = "p5.48xl-topo.xml",
+		.topology = NULL,
 		.default_dup_conns = 0,
 		.latency = 75.0,
 		.gdr_required = true,