-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
* 2003, Copyright, Hewlett-Packard Development Compnay, LP.
*
* Developed under the sponsorship of the U.S. Government
/*
* Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
* Use is subject to license terms.
+ *
+ * Copyright (c) 2012, Intel Corporation.
*/
#include <stdio.h>
#include "mpi.h"
/* lustre */
-#include <liblustre.h>
-#include <lustre/liblustreapi.h> /* for O_LOV_DELAY_CREATE */
+#include <lustre/lustreapi.h> /* for O_LOV_DELAY_CREATE */
#define CHECK_COUNT 10000
#define DISPLAY_COUNT (CHECK_COUNT * 10)
VERBOSE = 'V',
DEBUG = 'v',
HELP = 'h',
+ MNT = 'M',
+ MNTCOUNT = 'N',
+ MDTCOUNT = 'T',
};
struct option longOpts[] = {
{"verbose", 0, NULL, VERBOSE },
{"debug", 0, NULL, DEBUG },
{"help", 0, NULL, HELP },
- { 0, 0, NULL, 0 }
+ {"mdtcount", 1, NULL, MDTCOUNT },
+ {"mntcount", 1, NULL, MNTCOUNT },
+ {"mntfmt", 1, NULL, MNT },
+ { 0, 0, NULL, 0 }
};
int foo1, foo2;
int nfiles;
char filefmt[PATH_MAX];
char filename[PATH_MAX];
+char path[PATH_MAX];
int stripes = -1;
int begin;
int beginsave;
int verbose;
int debug;
struct stat statbuf;
+int mnt_count = -1;
+int mdt_count = 1;
+char *mntfmt;
#define dmesg if (debug) printf
#define DISPLAY_PROGRESS() { \
- if ((++nops % CHECK_COUNT) == 0 && verbose) { \
- curTime = time(0); \
+ if (verbose && (nops % CHECK_COUNT == 0)) { \
+ curTime = MPI_Wtime(); \
interval = curTime - lastTime; \
if (interval > DISPLAY_TIME || nops % DISPLAY_COUNT == 0) { \
- rate = (float)(nops - lastOps); \
- if (interval > 1) \
- rate /= (float)interval; \
- printf("Rank %d: %.2f %ss/sec %lu secs " \
- "(total: %d %ss %lu secs)\n", \
+ rate = (double)(nops - lastOps)/interval; \
+ printf("Rank %d: %.2f %ss/sec %.2f secs " \
+ "(total: %d %ss %.2f secs)\n", \
myrank, rate, cmd, interval, \
nops, cmd, curTime - startTime); \
lastOps = nops; \
" [ --dirfmt <str> ] [ --ndirs <num> ]\n"
" [ --filefmt <str> ] [ --stripes <num> ]\n"
" [ --random_order [--seed <num> | --seedfile <file>] ]\n"
- " [ --readdir_order ]\n";
+ " [ --readdir_order ] [ --mntfmt <str> ]\n"
+ " [ --mntcount <num> ] [ --mdtcount <num> ]\n";
static void
usage(FILE *stream, char *fmt, ...)
if ((*endptr != 0) || (iters <= 0)) {
fatal(0, "Invalid --iters value.\n");
}
- if (mode != LOOKUP && mode != OPEN && mode != STAT) {
+ if (mode != LOOKUP && mode != OPEN) {
usage(stderr, "--iters only makes sense with "
- "--lookup, --open, or --stat.\n");
+ "--lookup or --open.\n");
}
break;
case TIME:
break;
case RANDOM:
case READDIR:
- if (mode != LOOKUP && mode != OPEN && mode != STAT) {
+ if (mode != LOOKUP && mode != OPEN) {
fatal(0, "--%s can only be specified with "
- "--lookup, --open, or --stat.\n",
+ "--lookup, or --open.\n",
(char *)longOpts[index].name);
}
order = rc;
break;
case HELP:
usage(stdout, NULL);
+ break;
+ case MNT:
+ if (strlen(optarg) > (PATH_MAX - 16))
+ fatal(0, "--mnt too long\n");
+ mntfmt = optarg;
+ break;
+ case MNTCOUNT:
+ mnt_count = strtol(optarg, &endptr, 0);
+ if ((*endptr != 0) || (mnt_count <= 0)) {
+ fatal(0, "Invalid --mnt_count value %s.\n",
+ optarg);
+ }
+ break;
+ case MDTCOUNT:
+ mdt_count = strtol(optarg, &endptr, 0);
+ if ((*endptr != 0) || (mdt_count <= 0)) {
+ fatal(0, "Invalid --mdt_count value %s.\n",
+ optarg);
+ }
+ break;
default:
usage(stderr, "unrecognized option: '%c'.\n", optopt);
}
usage(stderr, "too many arguments %d >= %d.\n", optind, argc);
}
- if (mode == CREATE || mode == MKNOD || mode == UNLINK) {
+ if ((mnt_count != -1 && mntfmt == NULL) ||
+ (mnt_count == -1 && mntfmt != NULL)) {
+ usage(stderr, "mnt_count and mntfmt must be specified at the "
+ "same time\n");
+ }
+
+ if (mode == CREATE || mode == MKNOD || mode == UNLINK || mode == STAT) {
if (seconds != 0) {
if (nfiles == 0)
nfiles = INT_MAX;
usage(stderr, "--nfiles or --time must be specified "
"with %s.\n", cmd);
}
- } else if (mode == LOOKUP || mode == OPEN || mode == STAT) {
+ } else if (mode == LOOKUP || mode == OPEN) {
if (seconds != 0) {
if (iters == 0)
iters = INT_MAX;
if ((end > tmpend) || (end <= 0))
end -= dirthreads;
+ /* make sure mnt_count <= nthreads, otherwise it might div 0 in
+ * the following test */
+ if (mnt_count > nthreads)
+ mnt_count = nthreads;
+
begin += offset;
if (begin < 0)
begin = INT_MAX;
if (dirfmt == NULL) {
strcpy(dir, ".");
} else {
- sprintf(dir, dirfmt, dirnum);
-
- sprintf(mkdir_cmd, "/bin/mkdir -p %s", dir);
- #ifdef _LIGHTWEIGHT_KERNEL
- printf("NOTICE: not running system(%s)\n", mkdir_cmd);
- #else
- rc = system(mkdir_cmd);
- if (rc) {
- fatal(myrank, "'%s' failed.\n", mkdir_cmd);
- }
- #endif
+ int dir_len = 0;
+
+ if (mntfmt != NULL) {
+ sprintf(dir, mntfmt, (myrank / (nthreads/mnt_count)));
+ strcat(dir, "/");
+ dir_len = strlen(dir);
+ }
+ sprintf(dir + dir_len, dirfmt, dirnum);
+
+ if (mdt_count > 1) {
+ struct stat sb;
+ if (stat(dir, &sb) == 0) {
+ if (!S_ISDIR(sb.st_mode))
+ fatal(myrank, "'%s' is not dir\n", dir);
+ } else if (errno == ENOENT) {
+ sprintf(mkdir_cmd, "lfs mkdir -i %d %s",
+ myrank % mdt_count, dir);
+ } else {
+ fatal(myrank, "'%s' stat failed\n", dir);
+ }
+ } else {
+ sprintf(mkdir_cmd, "mkdir -p %s", dir);
+ }
+
+ dmesg("%d: %s\n", myrank, mkdir_cmd);
+#ifdef _LIGHTWEIGHT_KERNEL
+ printf("NOTICE: not running system(%s)\n", mkdir_cmd);
+#else
+ rc = system(mkdir_cmd);
+ if (rc)
+ fatal(myrank, "'%s' failed.\n", mkdir_cmd);
+#endif
rc = chdir(dir);
if (rc) {
int
main(int argc, char *argv[])
{
- int i, j, fd, rc, nops, lastOps, ag_ops;
- int ag_interval = 0;
- float rate, ag_rate;
- float avg_rate = 0;
- float effective_rate = 0;
- time_t startTime, lastTime, curTime, interval;
+ int i, j, fd, rc, nops, lastOps;
+ int ag_ops = 0;
+ double ag_interval = 0;
+ double ag_rate = 0;
+ double rate, avg_rate, effective_rate;
+ double startTime, curTime, lastTime, interval;
+ time_t timestamp;
char * file;
rc = MPI_Init(&argc, &argv);
process_args(argc, argv);
- startTime = time(0);
+ timestamp = time(0);
if ((myrank == 0) || debug) {
printf("%d: %s starting at %s",
- myrank, hostname, ctime(&startTime));
+ myrank, hostname, ctime(×tamp));
}
/* if we're not measuring creation rates then precreate
* the files we're operating on. */
- if ((mode != CREATE) && (mode != MKNOD) && !ignore) {
+ if ((mode != CREATE) && (mode != MKNOD) && !ignore &&
+ (mode != UNLINK || recreate)) {
/* create the files in reverse order. When we encounter
* a file that already exists, assume the remainder of
* the files exist to save time. The timed performance
dir, strerror(rc));
}
- startTime = time(0);
+ timestamp = time(0);
j = random() % nfiles;
dmesg("%d: %s initializing dir offset %u: %s",
- myrank, hostname, j, ctime(&startTime));
+ myrank, hostname, j, ctime(×tamp));
for (i = 0; i <= j; i++) {
if ((dir_entry = readdir(directory)) == NULL) {
}
}
- lastTime = time(0);
+ timestamp = time(0);
dmesg("%d: index %d, filename %s, offset %ld: "
"%s initialization complete: %s",
myrank, i, dir_entry->d_name, telldir(directory),
- hostname, ctime(&lastTime));
+ hostname, ctime(×tamp));
}
- rc = MPI_Barrier(MPI_COMM_WORLD);
- if (rc != MPI_SUCCESS)
- fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
-
if (seconds) {
act.sa_handler = sigalrm_handler;
(void)sigemptyset(&act.sa_mask);
alarm(seconds);
}
- startTime = lastTime = time(0);
+ rc = MPI_Barrier(MPI_COMM_WORLD);
+ if (rc != MPI_SUCCESS)
+ fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
+
+ startTime = lastTime = MPI_Wtime();
nops = lastOps = 0;
switch (mode) {
}
close(fd);
+ nops++;
DISPLAY_PROGRESS();
}
"error: %s\n", filename, strerror(rc));
}
+ nops++;
DISPLAY_PROGRESS();
}
break;
filename, strerror(rc));
}
+ nops++;
DISPLAY_PROGRESS();
}
break;
close(fd);
+ nops++;
DISPLAY_PROGRESS();
}
break;
case STAT:
- for (; nops < iters && !alarm_caught;) {
- rc = stat(file = next_file(), &statbuf);
+ for (; begin <= end && !alarm_caught; begin += dirthreads) {
+ sprintf(filename, filefmt, begin);
+ rc = stat(filename, &statbuf);
if (rc) {
if (((rc = errno) == EINTR) && alarm_caught)
break;
+ if (((rc = errno) == ENOENT) && ignore)
+ continue;
fatal(myrank, "stat(%s) error: %s\n",
- file, strerror(rc));
+ filename, strerror(rc));
}
+ nops++;
DISPLAY_PROGRESS();
}
break;
if (rc) {
if (((rc = errno) == EINTR) && alarm_caught)
break;
- if (((rc = errno) == ENOENT) && ignore)
- continue;
+ if ((rc = errno) == ENOENT) {
+ if (ignore)
+ continue;
+ /* no more files to unlink */
+ break;
+ }
fatal(myrank, "unlink(%s) error: %s\n",
filename, strerror(rc));
}
+ nops++;
DISPLAY_PROGRESS();
}
break;
}
- curTime = time(0);
+ rc = MPI_Barrier(MPI_COMM_WORLD);
+ if (rc != MPI_SUCCESS)
+ fatal(myrank, "prep MPI_Barrier failed: %d\n", rc);
+ curTime = MPI_Wtime();
interval = curTime - startTime;
- rate = (float)(nops);
- if (interval != 0)
- rate /= (float)interval;
+ rate = (double) (nops) / interval;
rc = MPI_Reduce(&nops, &ag_ops, 1, MPI_INT, MPI_SUM, 0,
MPI_COMM_WORLD);
fatal(myrank, "Failure in MPI_Reduce of total ops.\n");
}
- rc = MPI_Reduce(&interval, &ag_interval, 1, MPI_INT, MPI_SUM, 0,
+ rc = MPI_Reduce(&interval, &ag_interval, 1, MPI_DOUBLE, MPI_SUM, 0,
MPI_COMM_WORLD);
if (rc != MPI_SUCCESS) {
fatal(myrank, "Failure in MPI_Reduce of total interval.\n");
}
- rc = MPI_Reduce(&rate, &ag_rate, 1, MPI_FLOAT, MPI_SUM, 0,
+ rc = MPI_Reduce(&rate, &ag_rate, 1, MPI_DOUBLE, MPI_SUM, 0,
MPI_COMM_WORLD);
if (rc != MPI_SUCCESS) {
fatal(myrank, "Failure in MPI_Reduce of aggregated rate.\n");
}
if (myrank == 0) {
- curTime = time(0);
- interval = curTime - startTime;
- if (interval != 0)
- effective_rate = (float) ag_ops/ (float)interval;
-
- if (ag_interval > 0)
- avg_rate = (float)ag_ops/(float)ag_interval;
-
- printf("Rate: %.2f eff %.2f aggr %.2f client %ss/sec "
- "(total: %d threads %d %ss %d dirs %d threads/dir %lu secs)\n",
- effective_rate, ag_rate, avg_rate, cmd, nthreads, ag_ops, cmd,
- ndirs, dirthreads, interval);
+ curTime = MPI_Wtime();
+ interval = curTime - startTime;
+ effective_rate = (double) ag_ops / interval;
+ avg_rate = (double) ag_ops / ag_interval;
+
+ printf("Rate: %.2f eff %.2f aggr %.2f avg client %ss/sec "
+ "(total: %d threads %d %ss %d dirs %d threads/dir %.2f secs)\n",
+ effective_rate, ag_rate, avg_rate, cmd, nthreads, ag_ops,
+ cmd, ndirs, dirthreads, interval);
+ if (mode == UNLINK && !recreate && !ignore && ag_ops != nfiles)
+ printf("Warning: only unlinked %d files instead of %d"
+ "\n", ag_ops, nfiles);
}
if (recreate) {
}
}
- curTime = time(0);
+ timestamp = time(0);
if ((myrank == 0) || debug) {
printf("%d: %s finished at %s",
- myrank, hostname, ctime(&curTime));
+ myrank, hostname, ctime(×tamp));
}
MPI_Finalize();