#include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "util.h" #define NREQS 1 /* # of parallel I/O requests per process */ #define SZBUF (1ULL<<23) #define MYTIME_TOUSEC 1000000 #define MYTIME_TONSEC 1000000000 #define NROW 11 #define NCOL 4 #define NSAMPLES_DROP 0/*10*/ #define NSAMPLES_IO 2/*20*/ #define NSAMPLES_TOTAL 2/*20*/ #define NSAMPLES_INNER 1 #define Q(x) #x #define QUOTE(x) Q(x) char test_srcdir[PATH_MAX]; static inline double mytime() { return /*rdtsc_light()*/MPI_Wtime(); } struct aioreq { int rank; int status; struct aiocb *aiocbp; }; static void aio_sighandler(int sig, siginfo_t *si, void *ucontext) { if (si->si_code == SI_ASYNCIO) { //struct aioreq *aioreq = si->si_value.sival_ptr; //pr_debug("I/O completion signal received\n"); } } int my_aio_init(int nreqs, struct aioreq *iolist, struct aiocb *aiocblist, char *aiobufs[NREQS]) { int j; for (j = 0; j < nreqs; j++) { iolist[j].rank = j; iolist[j].aiocbp = &aiocblist[j]; iolist[j].aiocbp->aio_buf = aiobufs[j]; iolist[j].aiocbp->aio_nbytes = SZBUF; iolist[j].aiocbp->aio_reqprio = 0; iolist[j].aiocbp->aio_offset = 0; iolist[j].aiocbp->aio_sigevent.sigev_notify = SIGEV_SIGNAL; iolist[j].aiocbp->aio_sigevent.sigev_signo = SIGUSR1; iolist[j].aiocbp->aio_sigevent.sigev_value.sival_ptr = &iolist[j]; } return 0; } int my_aio_evict(int nreqs, char **fn) { int ret; int i; char cmd[PATH_MAX]; for (i = 0; i < NREQS; i++) { sprintf(cmd, "%s -e %s > /dev/null", QUOTE(VMTOUCH), fn[i]); ret = system(cmd); if (ret == -1) { pr_err("%s: error: system\n", __func__); goto out; } if (WEXITSTATUS(ret)) { pr_err("%s: error: system returned %d\n", __func__, WEXITSTATUS(ret)); ret = WEXITSTATUS(ret); goto out; } } ret = 0; out: return ret; } int my_aio_open(int nreqs, struct aioreq *iolist, char **fn) { int ret; int j; for (j = 0; j < NREQS; j++) { iolist[j].aiocbp->aio_fildes = open(fn[j], O_RDWR | O_CREAT | O_TRUNC, S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP | S_IROTH); if (iolist[j].aiocbp->aio_fildes == -1) { pr_err("%s: error: open %s: %s\n", __func__, fn[j], strerror(errno)); ret = 1; goto out; } } ret = 0; out: return ret; } int my_aio_check(int nreqs, char **fn, char **mem_data) { int ret; int i; FILE *fp[NREQS] = { 0 }; char *file_data[NREQS]; /* Check contents */ for (i = 0; i < nreqs; i++) { if (!(file_data[i] = malloc(SZBUF))) { pr_err("error: allocating data\n"); ret = -ENOMEM; goto out; } if (!(fp[i] = fopen(fn[i], "r+"))) { pr_err("error: fopen %s: %s\n", fn[i], strerror(errno)); ret = -errno; goto out; } if (fread(file_data[i], sizeof(char), SZBUF, fp[i]) != SZBUF) { pr_err("error: fread: %s\n", strerror(errno)); ret = -1; goto out; } fclose(fp[i]); if (memcmp((const char *)file_data[i], mem_data[i], SZBUF)) { pr_err("%s: file_data[%d] and mem_data[%d] doesn't match\n", __func__, i, i); ret = -1; goto out; } free(file_data[i]); } ret = 0; out: return ret; } void my_aio_close(int nreqs, struct aioreq *iolist) { int j; for (j = 0; j < NREQS; j++) { close(iolist[j].aiocbp->aio_fildes); iolist[j].aiocbp->aio_fildes = -1; } } int my_aio(int nreqs, struct aioreq *iolist, char **fn, long nsec_calc) { int ret; int i, j; /* Start async IO */ for (j = 0; j < NSAMPLES_INNER; j++) { int completion_count = 0; //pr_debug("debug: opening file\n"); if ((ret = my_aio_open(nreqs, iolist, fn)) == -1) { pr_err("%s: error: aio_read: %s\n", __func__, strerror(errno)); ret = -errno; goto out; } //pr_debug("debug: issuing write command\n"); for (j = 0; j < nreqs; j++) { /* Reset completion notice */ iolist[j].status = EINPROGRESS; if ((ret = aio_write(iolist[j].aiocbp)) == -1) { pr_err("%s: error: aio_read: %s\n", __func__, strerror(errno)); ret = -errno; goto out; } } /* Emulate calcuation phase */ ndelay(nsec_calc); /* Wait for completion of async IO */ //pr_debug("debug: waiting for completion\n"); while (completion_count != nreqs) { for (j = 0; j < nreqs; j++) { if (iolist[j].status != EINPROGRESS) { continue; } iolist[j].status = aio_error(iolist[j].aiocbp); switch (iolist[j].status) { case 0: /* Succeeded */ goto completed; case EINPROGRESS: break; case ECANCELED: pr_err("%s: error: aio is cancelled\n", __func__); goto completed; default: pr_err("%s: error: unexpected status: %d\n", __func__, iolist[j].status); goto completed; completed: completion_count++; break; } } } /* Check write amount */ for (j = 0; j < nreqs; j++) { ssize_t size; if ((size = aio_return(iolist[j].aiocbp)) != SZBUF) { pr_err("%s: Expected to have written %ld B but reported to have written %ld B\n", __func__, SZBUF, size); ret = -1; goto out; } } my_aio_close(nreqs, iolist); } ret = 0; out: return ret; } int measure(double *result, int nsamples, int nsamples_drop, int nreqs, struct aioreq *iolist, char **fn, char **aiobufs, long nsec_calc) { int ret; int i; double t_l, t_g, t_sum = 0; double start, end; for (i = 0; i < nsamples + nsamples_drop; i++) { #if 0 pr_debug("debug: evicting file cache\n"); if ((ret = my_aio_evict(nreqs, fn))) { pr_err("%s: error: my_aio_evict returned %d\n", __func__, ret); } #endif MPI_Barrier(MPI_COMM_WORLD); start = mytime(); if ((ret = my_aio(nreqs, iolist, fn, nsec_calc))) { pr_err("%s: error: my_aio_read returned %d\n", __func__, ret); } end = mytime(); MPI_Barrier(MPI_COMM_WORLD); /* Check contents */ if ((ret = my_aio_check(nreqs, fn, aiobufs))) { pr_err("%s: error: my_aio_check returned %d\n", __func__, ret); } if (i < nsamples_drop) { continue; } /* Take max */ t_l = end - start; MPI_Allreduce(&t_l, &t_g, 1, MPI_DOUBLE, MPI_MAX, MPI_COMM_WORLD); t_sum += t_g; } *result = t_sum / nsamples; ret = 0; out: return ret; } int main(int argc, char **argv) { int ret; int i, j, progress, l; int rank, nproc; int disable_syscall_intercept = 0, ppn = -1; struct aioreq *iolist; struct aiocb *aiocblist; struct sigaction sa; double t_io_ave, t_total_ave; double t_table[NROW][NCOL] = { 0 }; int opt; char *aiobufs[NREQS] = { 0 }; char **fn; opterr = 0; /* Don't print out error when not recognizing option character */ while ((opt = getopt(argc, argv, ":I:p:")) != -1) { switch (opt) { case 'I': disable_syscall_intercept = atoi(optarg); break; case 'p': ppn = atoi(optarg); break; case '?': pr_err("error: invalid option: -%c\n", optopt); ret = 1; goto out; case ':': pr_err("error: option -%c requires an argument\n", optopt); ret = 1; goto out; } } if (ppn == -1) { pr_err("error: specify -p \n"); ret = -EINVAL; goto out; } /* Initialize MPI */ MPI_Init(&argc, &argv); MPI_Comm_rank(MPI_COMM_WORLD, &rank); MPI_Comm_size(MPI_COMM_WORLD, &nproc); /* Show parameters */ if (rank == 0) { #pragma omp parallel { if (omp_get_thread_num() == 0) { printf("nproc: %d, ppn: %d, #threads: %d\n", nproc, ppn, omp_get_num_threads()); } } } /* Set verbosity */ //test_set_loglevel(TEST_LOGLEVEL_WARN); /* Initialize delay function */ ndelay_init(); /* Prepare file names */ #define TEST_SRCDIR "/work/gg10/e29005" sprintf(test_srcdir, "%s", /*TEST_SRCDIR*/dirname(argv[0])); if (!(fn = malloc(sizeof(char *) * NREQS))) { pr_err("error: allocating fn\n"); ret = -ENOMEM; goto out; } for (i = 0; i < NREQS; i++) { if (!(fn[i] = malloc(PATH_MAX))) { pr_err("error: allocating fn\n"); ret = -ENOMEM; goto out; } sprintf(fn[i], "%s/rank%d-number%d", test_srcdir, rank, i); if (rank == 0) pr_debug("debug: rank: %d, fn[%d]: %s\n", rank, i, fn[i]); } /* Allocate aio commands */ if (!(iolist = calloc(NREQS, sizeof(struct aioreq)))) { pr_err("%s: error: allocating iolist\n", __func__); ret = 1; goto out; } if (!(aiocblist = calloc(NREQS, sizeof(struct aiocb)))) { pr_err("%s: error: allocating aiocblist\n", __func__); ret = 1; goto out; } /* Prepare contents to be written */ for (i = 0; i < NREQS; i++) { aiobufs[i] = malloc(SZBUF); if (!aiobufs[i]) { pr_err("%s: error: allocating aiobufs\n", __func__); ret = 1; goto out; } for (j = 0; j < SZBUF; j++) { *(aiobufs[i] + j) = i + j + rank; } } /* Set signal handlers */ sa.sa_flags = SA_RESTART | SA_SIGINFO; sa.sa_sigaction = aio_sighandler; if (sigaction(SIGUSR1, &sa, NULL) == -1) { pr_err("%s: error: sigaction: %s\n", __func__, strerror(errno)); ret = 1; goto out; } /* Set aio parameters except fd and status */ if ((ret = my_aio_init(NREQS, iolist, aiocblist, aiobufs))) { pr_err("%s: error: my_aio_init returned %d\n", __func__, ret); goto out; } /* Measure IO only time */ //pr_debug("debug: measuring IO only time\n"); if ((ret = measure(&t_io_ave, NSAMPLES_IO, NSAMPLES_DROP, NREQS, iolist, fn, aiobufs, 0))) { pr_err("error: measure returned %d\n", ret); goto out; } if (rank == 0) { printf("t_io_ave: %.0f usec, %.0f MB/s per node\n", t_io_ave * MYTIME_TOUSEC, SZBUF * ppn / t_io_ave / 1000000); } /* Measure time with no progress, progress and no uti, progress and uti */ for (progress = 0; progress <= (disable_syscall_intercept ? 0 : 0); progress += 1) { /* Spawn helper thread onto compute CPUs with ignoring uti_attr */ if (progress == 1) { setenv("DISABLE_UTI", "1", 1); } /* Spawn helper thread onto dedicated CPUs with respecting uti_attr */ else if (progress == 2) { unsetenv("DISABLE_UTI"); } /* Measure with various calculation time */ for (l = 0; l <= 10; l += 2) { long nsec_calc = (t_io_ave * MYTIME_TONSEC * l) / 10; if ((ret = measure(&t_total_ave, NSAMPLES_TOTAL, NSAMPLES_DROP, NREQS, iolist, fn, aiobufs, nsec_calc))) { pr_err("error: measure returned %d\n", ret); goto out; } if (rank == 0) { if (l == 0) { pr_debug("progress=%d\n", progress); if (progress == 0) { pr_debug("calc\ttotal\n"); } else { pr_debug("total\n"); } } t_table[l][0] = nsec_calc * (MYTIME_TOUSEC / (double)MYTIME_TONSEC); if (progress == 0) { pr_debug("%.0f\t%.0f\n", nsec_calc * (MYTIME_TOUSEC / (double)MYTIME_TONSEC), t_total_ave * MYTIME_TOUSEC); t_table[l][progress + 1] = t_total_ave * MYTIME_TOUSEC; } else { pr_debug("%.0f\n", t_total_ave * MYTIME_TOUSEC); t_table[l][progress + 1] = t_total_ave * MYTIME_TOUSEC; } } } } if (rank == 0) { printf("calc,no prog,prog and no uti, prog and uti\n"); for (l = 0; l <= 10; l++) { for (i = 0; i < NCOL; i++) { if (i > 0) { printf(","); } printf("%.0f", t_table[l][i]); } printf("\n"); } } MPI_Barrier(MPI_COMM_WORLD); MPI_Finalize(); ret = 0; out: for (i = 0; i < NREQS; i++) { free(aiobufs[i]); } return ret; }