#include #include #include #include #include #include #include #include #include #include #include #include #include /* For SYS_xxx definitions */ #include #include "util.h" #include "fwq.h" //#define DEBUG #ifdef DEBUG #define dprintf printf #else #define dprintf {} #endif #if 1 #define BEGIN_EPOCH(win) do { MPI_Win_fence(0, win); } while (0) #define END_EPOCH(win) do { MPI_Win_fence(0, win); } while (0) #define BAR_EPOCH do { } while (0) #else #define BEGIN_EPOCH(win) do { MPI_Win_lock_all(0, win); } while (0) #define END_EPOCH(win) do { MPI_Win_unlock_all(win); } while (0) #define BAR_EPOCH do { MPI_Barrier(MPI_COMM_WORLD); } while (0) #endif static inline int on_same_node(int ppn, int me, int you) { return (me / ppn == you / ppn); } /* fence-accumulate-calc-fence */ void accumulate(int nproc, int ppn, int rank, double *wbuf, double *rbuf, int ndoubles, MPI_Win win, long calc_nsec) { int i, j; int r = 0, s = 0; int req = 0; BEGIN_EPOCH(win); for (i = 0; i < nproc; i++) { if (!on_same_node(ppn, rank, i)) { for (j = 0; j < ndoubles; j++) { #if 0 printf("i=%d,j=%d,rbuf=%f,wbuf=%f\n", i, j, rbuf[i * ndoubles + j], wbuf[i * ndoubles + j]); #endif MPI_Accumulate(rbuf + i * ndoubles + j, 1, MPI_DOUBLE, i, i * ndoubles + j, 1, MPI_DOUBLE, MPI_SUM, win); } } } fwq(calc_nsec); END_EPOCH(win); } static struct option options[] = { { .name = "ppn", .has_arg = required_argument, .flag = NULL, .val = 'P', }, /* end */ { NULL, 0, NULL, 0, }, }; int main(int argc, char **argv) { int rc; int actual; int ppn = -1; int nproc; int ndoubles = -1; int my_rank = -1, size = -1; int i, j; double *wbuf, *rbuf; MPI_Win win; struct timespec start, end; long t_fence_l, t_pure_l, t_overall_l; long t_fence, t_pure, t_overall; int opt; fwq_init(); while ((opt = getopt_long(argc, argv, "+d:P:", options, NULL)) != -1) { switch (opt) { case 'd': ndoubles = (1ULL << atoi(optarg)); break; case 'P': ppn = atoi(optarg); break; default: /* '?' */ printf("unknown option %c\n", optopt); exit(1); } } if (ndoubles == -1 || ppn == -1) { printf("specify ndoubles with -d and ppn with --ppn"); exit(1); } MPI_Init_thread(&argc, &argv, MPI_THREAD_MULTIPLE, &actual); if (actual != 3) { printf("ERROR: MPI_THREAD_MULTIPLE not available " "(level was set to %d)\n", actual); exit(1); } MPI_Comm_rank(MPI_COMM_WORLD, &my_rank); MPI_Comm_size(MPI_COMM_WORLD, &nproc); if (my_rank == 0) { printf("ndoubles=%d,nproc=%d\n", ndoubles, nproc); printf("nsec=%ld, nspw=%f\n", nsec, nspw); } /* write-to buffer */ wbuf = malloc(sizeof(double) * ndoubles * nproc); if (!wbuf) { printf("malloc failed"); goto fn_fail; } memset(wbuf, 0, sizeof(double) * ndoubles * nproc); /* read-from buffer */ rbuf = malloc(sizeof(double) * ndoubles * nproc); if (!rbuf) { printf("malloc failed"); goto fn_fail; } memset(rbuf, 0, sizeof(double) * ndoubles * nproc); if (rc = MPI_Win_create(wbuf, sizeof(double) * ndoubles * nproc, sizeof(double), MPI_INFO_NULL, MPI_COMM_WORLD, &win)) { printf("MPI_Win_create failed,rc=%d\n", rc); } print_cpu_last_executed_on("main"); for (i = 0; i < nproc; i++) { for (j = 0; j < ndoubles; j++) { wbuf[i * ndoubles + j] = i + 1 + j; rbuf[i * ndoubles + j] = (i + 1) * 2 + j; } } #if 0 for (i = 0; i < nproc; i++) { for (j = 0; j < ndoubles; j++) { printf("wbuf,proc=%d,j=%d,val=%f\n", i, j, wbuf[i * ndoubles + j]); printf("rbuf,proc=%d,j=%d,val=%f\n", i, j, rbuf[i * ndoubles + j]); } } #endif /* Measure fence-fence time */ MPI_Barrier(MPI_COMM_WORLD); #define NSKIP 5 #define NFENCE 30 for (i = 0; i < NFENCE + NSKIP; i++) { if (i == NSKIP) { clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start); } BEGIN_EPOCH(win); END_EPOCH(win); } BAR_EPOCH; clock_gettime(CLOCK_THREAD_CPUTIME_ID, &end); t_fence_l = DIFFNSEC(end, start) / NFENCE; //printf("t_fence (local): %ld usec\n", t_fence_l / 1000UL); MPI_Allreduce(&t_fence_l, &t_fence, 1, MPI_LONG, MPI_MAX, MPI_COMM_WORLD); if (my_rank == 0) printf("t_fence (max): %ld usec\n", t_fence / 1000UL); /* Measure fence-acc-fence time */ MPI_Barrier(MPI_COMM_WORLD); #define NPURE 30 for (i = 0; i < NPURE + NSKIP; i++) { if (i == NSKIP) { clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start); } accumulate(nproc, ppn, my_rank, wbuf, rbuf, ndoubles, win, 0); } BAR_EPOCH; clock_gettime(CLOCK_THREAD_CPUTIME_ID, &end); t_pure_l = DIFFNSEC(end, start) / NPURE; //printf("t_pure (local): %ld usec\n", t_pure_l / 1000UL); MPI_Allreduce(&t_pure_l, &t_pure, 1, MPI_LONG, MPI_MAX, MPI_COMM_WORLD); if (my_rank == 0) printf("t_pure (max): %ld usec\n", t_pure / 1000UL); #if 0 for (i = 0; i < nproc; i++) { for (j = 0; j < ndoubles; j++) { printf("wbuf,proc=%d,j=%d,val=%f\n", i, j, wbuf[i * ndoubles + j]); printf("rbuf,proc=%d,j=%d,val=%f\n", i, j, rbuf[i * ndoubles + j]); } } #endif /* Measure fenc-acc-calc-fence time */ MPI_Barrier(MPI_COMM_WORLD); #define NOVERALL 30 for (i = 0; i < NOVERALL + NSKIP; i++) { if (i == NSKIP) { clock_gettime(CLOCK_THREAD_CPUTIME_ID, &start); } accumulate(nproc, ppn, my_rank, wbuf, rbuf, ndoubles, win, t_pure - t_fence); } BAR_EPOCH; clock_gettime(CLOCK_THREAD_CPUTIME_ID, &end); t_overall_l = DIFFNSEC(end, start) / NOVERALL; //printf("t_overall (local): %ld usec\n", t_overall_l / 1000UL); MPI_Allreduce(&t_overall_l, &t_overall, 1, MPI_LONG, MPI_MAX, MPI_COMM_WORLD); if (my_rank == 0) printf("t_overall (max): %ld usec\n", t_overall / 1000UL); if (my_rank == 0) { long t_abs = (t_pure * 2) - t_overall; printf("overlap: %.2f %%\n", (t_abs * 100) / (double)t_pure); } fn_exit: MPI_Finalize(); return 0; fn_fail: goto fn_exit; }