IMB-MPI4: add mpi4 benchmark; bcast,reduce persistent benchmarks

Author: JuliaRS

Makefile

@@ -40,6 +40,10 @@ IMB-NBC:
 	make -C src_cpp -f Makefile TARGET=NBC
 	@cp src_cpp/IMB-NBC .
 
+IMB-MPI4:
+	make -C src_cpp -f Makefile TARGET=MPI4
+	@cp src_cpp/IMB-MPI4 .
+
 IMB-EXT:
 	make -C src_cpp -f Makefile TARGET=EXT
 	@cp src_cpp/IMB-EXT .

src_c/IMB_alltoall.c

@@ -349,5 +349,196 @@ Output variables:
     time[0] = t_pure;
 }
 
-#endif // NBC // MPI1
+#elif defined MPI4 // NBC
+
+/*************************************************************************/
+
+void IMB_alltoall_persist(struct comm_info* c_info,
+                        int size,
+                        struct iter_schedule* ITERATIONS,
+                        MODES RUN_MODE,
+                        double* time) {
+/*
+
+
+                      MPI-MPI4 benchmark kernel
+                      Benchmarks MPI_Alltoall_init
+
+
+
+Input variables:
+
+-c_info               (type struct comm_info*)
+                      Collection of all base data for MPI;
+                      see [1] for more information
+
+
+-size                 (type int)
+                      Basic message size in bytes
+
+-ITERATIONS           (type struct iter_schedule *)
+                      Repetition scheduling
+
+-RUN_MODE             (type MODES)
+
+
+Output variables:
+
+-time                 (type double*)
+                      Timing result per sample
+
+
+*/
+    int         i = 0;
+    Type_Size   s_size,
+                r_size;
+    int         s_num = 0,
+                r_num = 0;
+    MPI_Request request;
+    MPI_Status  status;
+    double      t_pure = 0.,
+                t_comp = 0.,
+                t_ovrlp = 0.;
+
+    /* GET SIZE OF DATA TYPE */
+    MPI_Type_size(c_info->s_data_type, &s_size);
+    MPI_Type_size(c_info->r_data_type, &r_size);
+
+    if ((s_size != 0) && (r_size != 0))
+    {
+        s_num = size / s_size;
+        r_num = size / r_size;
+    }
+
+    if (c_info->rank != -1) {
+        IMB_alltoall_pure_persist(c_info, size, ITERATIONS, RUN_MODE, &t_pure);
+
+        /* INITIALIZATION CALL */
+        IMB_cpu_exploit(t_pure, 1);
+
+        IMB_do_n_barriers (c_info->communicator, N_BARR);
+
+        // Create a persistent collective operation
+        MPI_ERRHAND(MPI_Alltoall_init((char*)c_info->s_buffer + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
+                                      s_num,
+                                      c_info->s_data_type,
+                                      (char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
+                                      r_num,
+                                      c_info->r_data_type,
+                                      c_info->communicator,
+                                      c_info->info,
+                                      &request));
+        for(i = 0; i < ITERATIONS->n_sample; i++)
+        {
+            t_ovrlp -= MPI_Wtime();
+            // Start the persistent request
+            MPI_ERRHAND(MPI_Start(&request));
+            t_comp -= MPI_Wtime();
+            IMB_cpu_exploit(t_pure, 0);
+            t_comp += MPI_Wtime();
+
+            MPI_Wait(&request, &status);
+            t_ovrlp += MPI_Wtime();
+
+            IMB_do_n_barriers(c_info->communicator, c_info->sync);
+        }
+        // Clean up
+        MPI_Request_free(&request);
+
+        t_ovrlp /= ITERATIONS->n_sample;
+        t_comp  /= ITERATIONS->n_sample;
+    }
+
+    time[0] = t_pure;
+    time[1] = t_ovrlp;
+    time[2] = t_comp;
+}
+
+/*************************************************************************/
+
+void IMB_alltoall_pure_persist(struct comm_info* c_info,
+                                int size,
+                                struct iter_schedule* ITERATIONS,
+                                MODES RUN_MODE,
+                                double* time) {
+/*
+
+
+                      MPI-MPI4 benchmark kernel
+                      Benchmarks MPI_Alltoall_init
+
+Input variables:
+
+-c_info               (type struct comm_info*)
+                      Collection of all base data for MPI;
+                      see [1] for more information
+
+
+-size                 (type int)
+                      Basic message size in bytes
+
+-ITERATIONS           (type struct iter_schedule *)
+                      Repetition scheduling
+
+-RUN_MODE             (type MODES)
+                      (only MPI-2 case: see [1])
+
+
+Output variables:
+
+-time                 (type double*)
+                      Timing result per sample
+
+
+*/
+    int         i = 0;
+    Type_Size   s_size,
+                r_size;
+    int         s_num = 0,
+                r_num = 0;
+    MPI_Request request;
+    MPI_Status  status;
+    double      t_pure = 0.;
+
+    /* GET SIZE OF DATA TYPE */
+    MPI_Type_size(c_info->s_data_type, &s_size);
+    MPI_Type_size(c_info->s_data_type, &r_size);
+
+    if ((s_size != 0) && (r_size != 0)) {
+        s_num = size / s_size;
+        r_num = size / r_size;
+    }
+
+    if(c_info->rank != -1) {
+        IMB_do_n_barriers(c_info->communicator, N_BARR);
+
+        // Create a persistent collective operation
+        MPI_ERRHAND(MPI_Alltoall_init((char*)c_info->s_buffer + i % ITERATIONS->s_cache_iter * ITERATIONS->s_offs,
+                                        s_num,
+                                        c_info->s_data_type,
+                                        (char*)c_info->r_buffer + i % ITERATIONS->r_cache_iter * ITERATIONS->r_offs,
+                                        r_num,
+                                        c_info->r_data_type,
+                                        c_info->communicator,
+                                        c_info->info,
+                                        &request));
+        for(i = 0; i < ITERATIONS->n_sample; i++) {
+            t_pure -= MPI_Wtime();
+            // Start the persistent request
+            MPI_ERRHAND(MPI_Start(&request));
+            MPI_Wait(&request, &status);
+            t_pure += MPI_Wtime();
+
+            IMB_do_n_barriers(c_info->communicator, c_info->sync);
+        }
+        // Clean up
+        MPI_Request_free(&request);
+
+        t_pure /= ITERATIONS->n_sample;
+    }
+
+    time[0] = t_pure;
+}
+
+#endif // NBC // MPI1 // MPI4