Add --bidir option to MPI sendrecv benchmark

mpi/examples/Makefile.am

@@ -466,7 +466,6 @@ benchs_sendrecv_bench_SOURCES += benchs/abstract_sendrecv_bench.c
 
 benchs_sendrecv_parallel_tasks_bench_SOURCES = benchs/sendrecv_parallel_tasks_bench.c
 benchs_sendrecv_parallel_tasks_bench_SOURCES += benchs/bench_helper.c
-benchs_sendrecv_parallel_tasks_bench_SOURCES += benchs/abstract_sendrecv_bench.c
 
 benchs_burst_SOURCES = benchs/burst.c
 benchs_burst_SOURCES += benchs/burst_helper.c

mpi/examples/benchs/abstract_sendrecv_bench.c

@@ -19,7 +19,7 @@
 
 
 
-void sendrecv_bench(int mpi_rank, starpu_pthread_barrier_t* thread_barrier)
+void sendrecv_bench(int mpi_rank, starpu_pthread_barrier_t* thread_barrier, int bidir)
 {
 	uint64_t iterations = LOOPS_DEFAULT;
 	uint64_t s = 0;
@@ -62,6 +62,7 @@ void sendrecv_bench(int mpi_rank, starpu_pthread_barrier_t* thread_barrier)
 	float* vector_recv = NULL;
 	double t1, t2, global_tstart, global_tend;
 	double* lats = malloc(sizeof(double) * LOOPS_DEFAULT);
+	starpu_mpi_req send_req, recv_req;
 
 	if (thread_barrier != NULL)
 	{
@@ -88,18 +89,38 @@ void sendrecv_bench(int mpi_rank, starpu_pthread_barrier_t* thread_barrier)
 			if (mpi_rank == 0)
 			{
 				t1 = starpu_timing_now();
-				starpu_mpi_send(handle_send, 1, 0, MPI_COMM_WORLD);
-				starpu_mpi_recv(handle_recv, 1, 1, MPI_COMM_WORLD, NULL);
+				if (bidir)
+				{
+					starpu_mpi_isend(handle_send, &send_req, 1, 0, MPI_COMM_WORLD);
+					starpu_mpi_irecv(handle_recv, &recv_req, 1, 1, MPI_COMM_WORLD);
+					starpu_mpi_wait(&send_req, MPI_STATUS_IGNORE);
+					starpu_mpi_wait(&recv_req, MPI_STATUS_IGNORE);
+				}
+				else
+				{
+					starpu_mpi_send(handle_send, 1, 0, MPI_COMM_WORLD);
+					starpu_mpi_recv(handle_recv, 1, 1, MPI_COMM_WORLD, NULL);
+				}
 				t2 = starpu_timing_now();
 
-				const double t = (t2 -t1) / 2;
+				const double t = (t2 - t1) / 2;
 
 				lats[j] = t;
 			}
 			else
 			{
-				starpu_mpi_recv(handle_recv, 0, 0, MPI_COMM_WORLD, NULL);
-				starpu_mpi_send(handle_send, 0, 1, MPI_COMM_WORLD);
+				if (bidir)
+				{
+					starpu_mpi_irecv(handle_recv, &recv_req, 0, 0, MPI_COMM_WORLD);
+					starpu_mpi_isend(handle_send, &send_req, 0, 1, MPI_COMM_WORLD);
+					starpu_mpi_wait(&recv_req, MPI_STATUS_IGNORE);
+					starpu_mpi_wait(&send_req, MPI_STATUS_IGNORE);
+				}
+				else
+				{
+					starpu_mpi_recv(handle_recv, 0, 0, MPI_COMM_WORLD, NULL);
+					starpu_mpi_send(handle_send, 0, 1, MPI_COMM_WORLD);
+				}
 			}
 
 			starpu_mpi_barrier(MPI_COMM_WORLD);

mpi/examples/benchs/abstract_sendrecv_bench.h

@@ -17,4 +17,4 @@
 #include <starpu.h>
 
 
-void sendrecv_bench(int mpi_rank, starpu_pthread_barrier_t* thread_barrier);
+void sendrecv_bench(int mpi_rank, starpu_pthread_barrier_t* thread_barrier, int bidir);

mpi/examples/benchs/sendrecv_bench.c

@@ -14,8 +14,8 @@
  * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  */
 /*
- * Basic send receive benchmark.
- * Inspired a lot from NewMadeleine examples/benchmarks/nm_bench_sendrecv.c
+ * Send-receive benchmark where peers exchange data at the same time.
+ * Inspired a lot from NewMadeleine examples/benchmarks/nm_bench_bidir.c
  */
 
 #include <starpu_mpi.h>
@@ -23,11 +23,22 @@
 #include "abstract_sendrecv_bench.h"
 
 
+static inline void man()
+{
+	fprintf(stderr, "Options:\n");
+	fprintf(stderr, "\t-h --help   display this help\n");
+	fprintf(stderr, "\t-p          pause workers during benchmark\n");
+	fprintf(stderr, "\t--bidir     full-duplex communications\n");
+	exit(EXIT_SUCCESS);
+}
+
+
 int main(int argc, char **argv)
 {
 	int ret, rank, worldsize;
 	int pause_workers = 0;
 	int i = 0;
+	int bidir = 0;
 
 
 	for (i = 1; i < argc; i++)
@@ -39,15 +50,17 @@ int main(int argc, char **argv)
 		}
 		else if (strcmp(argv[i], "-h") == 0 || strcmp(argv[i], "--help") == 0)
 		{
-			fprintf(stderr, "Options:\n");
-			fprintf(stderr, "\t-h --help   display this help\n");
-			fprintf(stderr, "\t-p          pause workers during benchmark\n");
-			exit(EXIT_SUCCESS);
+			man();
+		}
+		if (strcmp(argv[i], "--bidir") == 0)
+		{
+			bidir = 1;
+			printf("Communications will be full-duplex.\n");
 		}
 		else
 		{
 			fprintf(stderr,"Unrecognized option %s\n", argv[i]);
-			exit(EXIT_FAILURE);
+			man();
 		}
 	}
 
@@ -75,7 +88,7 @@ int main(int argc, char **argv)
 		starpu_pause();
 	}
 
-	sendrecv_bench(rank, NULL);
+	sendrecv_bench(rank, NULL, bidir);
 
 	if (pause_workers)
 	{

mpi/examples/benchs/sendrecv_gemm_bench.c

@@ -56,7 +56,7 @@ static void* comm_thread_func(void* arg)
 		fprintf(stderr, "[%s] No core was available for the comm thread. You should increase STARPU_RESERVE_NCPU or decrease STARPU_NCPU\n", hostname);
 	}
 
-	sendrecv_bench(mpi_rank, &thread_barrier);
+	sendrecv_bench(mpi_rank, &thread_barrier, /* half-duplex communications */ 0);
 
 	return NULL;
 }

mpi/examples/benchs/sendrecv_parallel_tasks_bench.c

@@ -34,7 +34,6 @@
 #include <starpu_mpi.h>
 #include "helper.h"
 #include "bench_helper.h"
-#include "abstract_sendrecv_bench.h"
 
 #define NB_WARMUP_PINGPONGS 10