1 /* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil ; -*- */
3 * (C) 2013 by Argonne National Laboratory.
4 * See COPYRIGHT in top-level directory.
12 #define MAX_DATA_SIZE (1024*128*16)
13 #define MAX_NUM_ITERATIONS (8192*4)
14 #define MIN_NUM_ITERATIONS 8
15 #define NUM_WARMUP_ITER 1
17 const int verbose = 0;
20 void run_test(int lock_mode, int lock_assert)
22 int nproc, test_iter, target_rank, data_size;
26 MPI_Comm_rank(MPI_COMM_WORLD, &rank);
27 MPI_Comm_size(MPI_COMM_WORLD, &nproc);
29 if (rank == 0 && verbose) {
30 printf("Starting one-sided contiguous performance test with %d processes\n", nproc);
32 printf("Synchronization mode: ");
35 case MPI_LOCK_EXCLUSIVE:
36 printf("Exclusive lock");
39 printf("Shared lock");
42 printf("Unknown lock");
46 if (lock_assert & MPI_MODE_NOCHECK)
47 printf(", MPI_MODE_NOCHECK");
52 MPI_Alloc_mem(MAX_DATA_SIZE, MPI_INFO_NULL, &buf);
53 MPI_Alloc_mem(MAX_DATA_SIZE, MPI_INFO_NULL, &win_buf);
54 memset(buf, rank, MAX_DATA_SIZE);
55 memset(win_buf, rank, MAX_DATA_SIZE);
56 MPI_Win_create(win_buf, MAX_DATA_SIZE, 1, MPI_INFO_NULL, MPI_COMM_WORLD, &win);
58 if (rank == 0 && verbose)
59 printf("%12s %12s %12s %12s %12s %12s %12s %12s\n", "Trg. Rank", "Xfer Size",
60 "Get (usec)", "Put (usec)", "Acc (usec)",
61 "Get (MiB/s)", "Put (MiB/s)", "Acc (MiB/s)");
63 for (target_rank = 0; rank == 0 && target_rank < nproc; target_rank++) {
64 for (data_size = sizeof(double); data_size <= MAX_DATA_SIZE; data_size *= 2) {
65 double t_get, t_put, t_acc;
66 int num_iter = MAX_NUM_ITERATIONS;
68 /* Scale the number of iterations by log_2 of the data size, so
69 * that we run each test for a reasonable amount of time. */
71 int t = data_size, my_log2 = 0;
75 num_iter = (num_iter / my_log2 < MIN_NUM_ITERATIONS) ?
76 MIN_NUM_ITERATIONS : num_iter / my_log2;
79 for (test_iter = 0; test_iter < num_iter + NUM_WARMUP_ITER; test_iter++) {
80 if (test_iter == NUM_WARMUP_ITER)
83 MPI_Win_lock(lock_mode, target_rank, lock_assert, win);
84 MPI_Get(buf, data_size, MPI_BYTE, target_rank, 0, data_size, MPI_BYTE, win);
85 MPI_Win_unlock(target_rank, win);
87 t_get = (MPI_Wtime() - t_get) / num_iter;
89 for (test_iter = 0; test_iter < num_iter + NUM_WARMUP_ITER; test_iter++) {
90 if (test_iter == NUM_WARMUP_ITER)
93 MPI_Win_lock(lock_mode, target_rank, lock_assert, win);
94 MPI_Put(buf, data_size, MPI_BYTE, target_rank, 0, data_size, MPI_BYTE, win);
95 MPI_Win_unlock(target_rank, win);
97 t_put = (MPI_Wtime() - t_put) / num_iter;
99 for (test_iter = 0; test_iter < num_iter + NUM_WARMUP_ITER; test_iter++) {
100 if (test_iter == NUM_WARMUP_ITER)
103 MPI_Win_lock(lock_mode, target_rank, lock_assert, win);
104 MPI_Accumulate(buf, data_size / sizeof(int), MPI_INT, target_rank,
105 0, data_size / sizeof(int), MPI_INT, MPI_SUM, win);
106 MPI_Win_unlock(target_rank, win);
108 t_acc = (MPI_Wtime() - t_acc) / num_iter;
110 if (rank == 0 && verbose)
111 printf("%12d %12d %12.3f %12.3f %12.3f %12.3f %12.3f %12.3f\n", target_rank,
112 data_size, t_get * 1.0e6, t_put * 1.0e6, t_acc * 1.0e6,
113 data_size / (1024.0 * 1024.0) / t_get, data_size / (1024.0 * 1024.0) / t_put,
114 data_size / (1024.0 * 1024.0) / t_acc);
118 MPI_Barrier(MPI_COMM_WORLD);
121 MPI_Free_mem(win_buf);
125 int main(int argc, char **argv)
127 MPI_Init(&argc, &argv);
129 run_test(MPI_LOCK_EXCLUSIVE, 0);
130 run_test(MPI_LOCK_EXCLUSIVE, MPI_MODE_NOCHECK);
131 run_test(MPI_LOCK_SHARED, 0);
132 run_test(MPI_LOCK_SHARED, MPI_MODE_NOCHECK);
137 printf(" No Errors\n");