int
main(int argc, char *argv[])
{
/* Check the rank, not using MPI. It wants to use MPI4PY on the
master only. Workers are simply to serve spawning. Use
Open-MPI specific environment variables. */
int vpid;
int nprocs;
{
char gomi[4];
char *e0 = "OMPI_MCA_orte_ess_vpid";
char *s0 = getenv(e0);
if (s0 == 0) {
fprintf(stderr, ("Environment variable %s needs to be set;"
" Seems not Open-MPI.\n"), e0);
fflush(0);
abort();
}
int v0;
int cc0 = sscanf(s0, "%d%c", &v0, gomi);
if (cc0 != 1) {
fprintf(stderr, "Bad environment variable %s (%s).\n", e0, s0);
fflush(0);
abort();
}
if (v0 < 0) {
fprintf(stderr, "Bad environment variable %s (%s).\n", e0, s0);
fflush(0);
abort();
}
char *e1 = "OMPI_MCA_orte_ess_num_procs";
char *s1 = getenv(e1);
if (s1 == 0) {
fprintf(stderr, ("Environment variable %s needs to be set;"
" Seems not Open-MPI.\n"), e1);
fflush(0);
abort();
}
int v1;
int cc1 = sscanf(s1, "%d%c", &v1, gomi);
if (cc1 != 1) {
fprintf(stderr, "Bad environment variable %s (%s).\n", e1, s1);
fflush(0);
abort();
}
if (v1 < 0) {
fprintf(stderr, "Bad environment variable %s (%s).\n", e1, s1);
fflush(0);
abort();
}
if (v0 >= v1) {
fprintf(stderr, ("Environment variables %s=%d and %s=%d"
" have bad values.\n"), e0, v0, e1, v1);
fflush(0);
abort();
}
vpid = v0;
nprocs = v1;
}
/* Start processing. */
int verbosity;
int xargc;
char **xargv;
if (argc >= 3 && strncasecmp(argv[1], "-v", 2) == 0) {
if (argv[1][2] == 0) {
verbosity = 2;
} else if (argv[1][2] == '0') {
verbosity = 0;
} else if (argv[1][2] == '1') {
verbosity = 1;
} else if (argv[1][2] == '2') {
verbosity = 2;
} else if (argv[1][2] == '3') {
verbosity = 3;
} else {
verbosity = 2;
}
xargc = (argc - 2);
xargv = &argv[2];
} else {
verbosity = -1;
xargc = (argc - 1);
xargv = &argv[1];
}
int master = (nprocs - 1);
if (vpid == master) {
/* MASTER */
if (xargc < 1) {
fprintf(stderr, "USAGE: %s python-script-file\n", argv[0]);
fflush(0);
int nprocs, rank, thlv;
MPI_Init_thread(&argc, &argv, MPI_THREAD_SERIALIZED, &thlv);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Abort(MPI_COMM_WORLD, 1);
exit(1);
}
char *cmd = xargv[0];
FILE *f = fopen(cmd, "r");
if (f == 0) {
char ee[80];
snprintf(ee, sizeof(ee), "fopen(%s): %s\n",
cmd, strerror(errno));
fprintf(stderr, ee);
fprintf(stderr, "USAGE: %s python-script-file\n", argv[0]);
fflush(0);
int nprocs, rank, thlv;
MPI_Init_thread(&argc, &argv, MPI_THREAD_SERIALIZED, &thlv);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Abort(MPI_COMM_WORLD, 1);
exit(1);
}
Py_SetProgramName(argv[0]);
Py_Initialize();
PyRun_SimpleFile(f, cmd);
Py_Finalize();
return 0;
} else {
/* WORKERS */
/* The actions of workers should exactly match with the master
in Python. It assumes MPI4PY itself only does local MPI
operations. */
int cc;
int nprocs, rank, thlv;
MPI_Init_thread(&argc, &argv, MPI_THREAD_SERIALIZED, &thlv);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
kmr_init();
#pragma omp parallel
{
sleep(0);
}
KMR *mr = kmr_create_context(MPI_COMM_WORLD, MPI_INFO_NULL, 0);
assert(mr != 0);
mr->trace_map_spawn = (verbosity > 0);
MPI_Comm splitcomms[4];
cc = kmr_split_swf_lanes(mr, splitcomms, master, 0, 1);
assert(cc == MPI_SUCCESS);
cc = kmr_init_swf(mr, splitcomms, master);
assert(cc == MPI_SUCCESS);
if (verbosity != -1) {
kmr_set_swf_verbosity(mr, verbosity);
}
cc = kmr_detach_swf_workers(mr);
assert(cc == MPI_SUCCESS);
/* NEVER COMES HERE. */
return 0;
}
}
static int
gen_cmdkvs(const struct kmr_kv_box kv,
const KMR_KVS *kvi, KMR_KVS *kvo, void *p, long i_)
{
char *cmd1 = "maxprocs=1 /bin/sleep 1";
char *cmd2 = "maxprocs=1 /bin/sleep 5";
char *cmd3 = "maxprocs=1 /bin/sleep 10";
int vlen = (int)strlen(cmd2) + 1;
for (int i = 0; i < NUM_COMMANDS; i++) {
char *cmd = NULL;
if (i % 3 == 0) {
cmd = cmd1;
} else if (i % 3 == 1) {
cmd = cmd2;
} else {
cmd = cmd3;
}
struct kmr_kv_box nkv = { .klen = sizeof(long),
.vlen = vlen * (int)sizeof(char),
.k.i = i,
.v.p = (void *)cmd };
kmr_add_kv(kvo, nkv);
}
return MPI_SUCCESS;
}
static int
output_result(const struct kmr_kv_box kv,
const KMR_KVS *kvi, KMR_KVS *kvo, void *p, long i_)
{
struct kmr_kv_box nkv = { .klen = sizeof(long),
.vlen = sizeof(long),
.k.i = kv.k.i,
.v.i = kv.k.i };
kmr_add_kv(kvo, nkv);
return MPI_SUCCESS;
}
int
main(int argc, char **argv)
{
MPI_Init(&argc, &argv);
kmr_init();
KMR *mr = kmr_create_context(MPI_COMM_WORLD, MPI_INFO_NULL, 0);
int rank;
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
fprintf(stderr, "Start\n");
}
MPI_Barrier(MPI_COMM_WORLD);
KMR_KVS *kvs_commands = kmr_create_kvs(mr, KMR_KV_INTEGER, KMR_KV_OPAQUE);
int ret = kmr_map_once(kvs_commands, 0, kmr_noopt, 1, gen_cmdkvs);
if (ret != MPI_SUCCESS) {
MPI_Abort(MPI_COMM_WORLD, 1);
}
kmr_dump_kvs(kvs_commands, 1);
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
fprintf(stderr, "MAP_ONCE DONE\n");
}
MPI_Barrier(MPI_COMM_WORLD);
KMR_KVS *kvs_runcmds = kmr_create_kvs(mr, KMR_KV_INTEGER, KMR_KV_OPAQUE);
ret = kmr_shuffle(kvs_commands, kvs_runcmds, kmr_noopt);
if (ret != MPI_SUCCESS) {
MPI_Abort(MPI_COMM_WORLD, 1);
}
kmr_dump_kvs(kvs_runcmds, 1);
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
fprintf(stderr, "SHUFFLE DONE\n");
}
MPI_Barrier(MPI_COMM_WORLD);
KMR_KVS *kvs_results = kmr_create_kvs(mr, KMR_KV_INTEGER, KMR_KV_INTEGER);
struct kmr_spawn_option sopt_sepsp = { .separator_space = 1 };
ret = kmr_map_serial_processes(kvs_runcmds, kvs_results, 0, MPI_INFO_NULL,
sopt_sepsp, output_result);
kmr_dump_kvs(kvs_results, 1);
kmr_free_kvs(kvs_results);
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
fprintf(stderr, "MAP_SPAWN DONE\n");
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
fprintf(stderr, "Finish\n");
}
kmr_free_context(mr);
kmr_fin();
MPI_Finalize();
}
static double
wtime()
{
static struct timeval tv0 = {.tv_sec = 0};
struct timeval tv;
int cc;
cc = gettimeofday(&tv, 0);
assert(cc == 0);
if (tv0.tv_sec == 0) {
tv0 = tv;
assert(tv0.tv_sec != 0);
}
double dt = ((double)(tv.tv_sec - tv0.tv_sec)
+ ((double)(tv.tv_usec - tv0.tv_usec) * 1e-6));
return dt;
}
/* Puts 200 key-value pairs to output KVO. It is a map-function. It
runs only on rank0. Inputs (KV0 and KVS0) are dummy. */
static int
addkeysfn(const struct kmr_kv_box kv0,
const KMR_KVS *kvs0, KMR_KVS *kvo, void *p, const long ind)
{
assert(kvs0 == 0 && kv0.klen == 0 && kv0.vlen == 0 && kvo != 0);
char k[80];
char v[80];
int cc;
for (int i = 0; i < 200; i++) {
snprintf(k, 80, "key%d", i);
snprintf(v, 80, "value%d", i);
struct kmr_kv_box kv = {
.klen = (int)(strlen(k) + 1),
.vlen = (int)(strlen(v) + 1),
.k.p = k,
.v.p = v
};
cc = kmr_add_kv(kvo, kv);
assert(cc == MPI_SUCCESS);
}
return MPI_SUCCESS;
}
static int
replacevaluefn(const struct kmr_kv_box kv0,
const KMR_KVS *kvs0, KMR_KVS *kvo, void *p,
const long i)
{
assert(kvs0 != 0 && kvo != 0);
int cc, x;
char gomi;
cc = sscanf((&((char *)kv0.k.p)[3]), "%d%c", &x, &gomi);
assert(cc == 1);
char v[80];
snprintf(v, 10, "newvalue%d", x);
struct kmr_kv_box kv = {.klen = kv0.klen,
.vlen = (int)(strlen(v) + 1),
.k.p = kv0.k.p,
.v.p = v
};
cc = kmr_add_kv(kvo, kv);
assert(cc == MPI_SUCCESS);
return MPI_SUCCESS;
}
static int
emptyreducefn(const struct kmr_kv_box kv[], const long n,
const KMR_KVS *kvs, KMR_KVS *kvo, void *p)
{
return MPI_SUCCESS;
}
/* Do KMR operations many times. */
static void
simple0(int nprocs, int rank)
{
int cc;
KMR *mr = kmr_create_context(MPI_COMM_WORLD, MPI_INFO_NULL, 0);
double t0, t1;
t0 = wtime();
for (int i = 0; i < 10000; i++) {
/* Check timeout. */
t1 = wtime();
KMR_KVS *to0 = kmr_create_kvs(mr, KMR_KV_INTEGER, KMR_KV_INTEGER);
if (rank == 0) {
struct kmr_kv_box kv = {
.klen = (int)sizeof(long),
.vlen = (int)sizeof(long),
.k.i = 0,
.v.i = ((t1 - t0) > 20.0)
};
cc = kmr_add_kv(to0, kv);
assert(cc == MPI_SUCCESS);
}
cc = kmr_add_kv_done(to0);
assert(cc == MPI_SUCCESS);
KMR_KVS *to1 = kmr_create_kvs(mr, KMR_KV_INTEGER, KMR_KV_INTEGER);
cc = kmr_replicate(to0, to1, kmr_noopt);
assert(cc == MPI_SUCCESS);
struct kmr_kv_box tok = {.klen = (int)sizeof(long), .k.p = 0,
.vlen = 0, .v.p = 0};
struct kmr_kv_box tov;
cc = kmr_find_key(to1, tok, &tov);
assert(cc == MPI_SUCCESS);
cc = kmr_free_kvs(to1);
assert(cc == MPI_SUCCESS);
if (tov.v.i) {
if (rank == 0) {
printf("loops %d\n", i);
}
break;
}
/* Put some pairs. */
KMR_KVS *kvs0 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_map_on_rank_zero(kvs0, 0, kmr_noopt, addkeysfn);
assert(cc == MPI_SUCCESS);
/* Replicate pairs to all ranks. */
KMR_KVS *kvs1 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_replicate(kvs0, kvs1, kmr_noopt);
assert(cc == MPI_SUCCESS);
/* Map pairs. */
KMR_KVS *kvs2 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_map(kvs1, kvs2, 0, kmr_noopt, replacevaluefn);
assert(cc == MPI_SUCCESS);
/* Collect pairs by theirs keys. */
KMR_KVS *kvs3 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_shuffle(kvs2, kvs3, kmr_noopt);
assert(cc == MPI_SUCCESS);
/* Reduce collected pairs. */
KMR_KVS *kvs4 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
cc = kmr_reduce(kvs3, kvs4, 0, kmr_noopt, emptyreducefn);
assert(cc == MPI_SUCCESS);
cc = kmr_free_kvs(kvs4);
assert(cc == MPI_SUCCESS);
}
cc = kmr_free_context(mr);
assert(cc == MPI_SUCCESS);
}
int
main(int argc, char *argv[])
{
char cmd[256];
int pid = getpid();
int N = 8;
int nprocs, rank, thlv;
/*MPI_Init(&argc, &argv);*/
MPI_Init_thread(&argc, &argv, MPI_THREAD_SERIALIZED, &thlv);
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
kmr_init();
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {printf("Check leakage by observing heap size.\n");}
if (rank == 0) {printf("Watch VSZ changes (loops %d times)...\n", N);}
if (rank == 0) {printf("(Each loop will take approx. 20 sec).\n");}
fflush(0);
usleep(50 * 1000);
MPI_Barrier(MPI_COMM_WORLD);
for (int i = 0; i < N; i++) {
simple0(nprocs, rank);
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) {
snprintf(cmd, sizeof(cmd), "ps l %d", pid);
system(cmd);
}
fflush(0);
}
MPI_Barrier(MPI_COMM_WORLD);
if (rank == 0) printf("OK\n");
fflush(0);
kmr_fin();
MPI_Finalize();
return 0;
}
static int
add_initial_data(const struct kmr_kv_box kv,
const KMR_KVS *kvi, KMR_KVS *kvo, void *p, long i_)
{
common_t *common = (common_t *)p;
char filename[FILENAME_LEN];
create_file(common->rank, common->iteration, common->file_size,
filename, FILENAME_LEN);
common->val_count = IO_COUNT * common->file_size;
struct kmr_kv_box nkv = { .klen = sizeof(char) * (strlen(common->key) + 1),
.k.p = common->key,
.vlen = sizeof(char) * (strlen(filename) + 1),
.v.p = (void *)filename };
kmr_add_kv(kvo, nkv);
return MPI_SUCCESS;
}
static int
increment_in_file_value(const struct kmr_kv_box kv,
const KMR_KVS *kvi, KMR_KVS *kvo, void *p, long i_)
{
common_t *common = (common_t *)p;
char *infile = (char *)kv.v.p;
char outfile[FILENAME_LEN];
snprintf(outfile, FILENAME_LEN, "./%06d-%02d.dat", common->rank,
common->iteration + 1);
FILE *ifp = fopen(infile, "r");
FILE *ofp = fopen(outfile, "w+");
assert(ifp != 0 && ofp != 0);
/* read/write 1MB at once */
long *buf = (long *)malloc(sizeof(long) * IO_COUNT);
for (int i = 0; i < common->file_size; i++) {
size_t cc = fread(buf, sizeof(long), IO_COUNT, ifp);
assert(cc == IO_COUNT);
for (int j = 0; j < IO_COUNT; j++) {
buf[j] += 1;
}
cc = fwrite(buf, sizeof(long), IO_COUNT, ofp);
assert(cc == IO_COUNT);
}
free(buf);
fclose(ofp);
struct kmr_kv_box nkv = { .klen = sizeof(char) * (strlen(common->key) + 1),
.k.p = common->key,
.vlen = sizeof(char) * (strlen(outfile) + 1),
.v.p = (void *)outfile };
kmr_add_kv(kvo, nkv);
#ifdef DEBUG
fseek(ifp, 0, SEEK_SET);
long val;
fread(&val, sizeof(long), 1, ifp);
fprintf(stderr, "Rank[%d]: process key[%s]-val[%ld]\n",
common->rank, (char *)kv.k.p, val);
#endif
fclose(ifp);
delete_file(common->rank, common->iteration);
return MPI_SUCCESS;
}
int
main(int argc, char **argv)
{
int thlv;
MPI_Init_thread(&argc, &argv, MPI_THREAD_SERIALIZED, &thlv);
int nprocs, rank, task_nprocs;
MPI_Comm_size(MPI_COMM_WORLD, &nprocs);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
check_nprocs(nprocs, rank, &task_nprocs);
kmr_init();
KMR *mr = kmr_create_context(MPI_COMM_WORLD, MPI_INFO_NULL, 0);
mr->verbosity = 5;
mr->trace_map_mp = 1;
char even_key[KEY_LEN];
char odd_key[KEY_LEN];
snprintf(even_key, KEY_LEN, "even%06d", (rank / task_nprocs + 1));
snprintf(odd_key, KEY_LEN, "odd%06d", (rank % task_nprocs + 1));
common_t common0;
common0.key = even_key;
parse_param_file(argc, argv, &(common0.file_size));
common0.rank = rank;
common0.iteration = 0;
KMR_KVS *kvs0 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
kmr_map_once(kvs0, &common0, kmr_noopt, 0, add_initial_data);
double itr_times[ITERATIONS];
for (int i = 0; i < ITERATIONS; i++) {
common0.key = (i % 2 == 0)? odd_key : even_key;
common0.iteration = i;
KMR_KVS *kvs1 = kmr_create_kvs(mr, KMR_KV_OPAQUE, KMR_KV_OPAQUE);
struct timeval ts;
measure_time(&ts);
kmr_map_multiprocess_by_key(kvs0, kvs1, &common0, kmr_noopt, rank,
increment_in_file_value);
struct timeval te;
measure_time(&te);
itr_times[i] = calc_time_diff(&ts, &te);
kvs0 = kvs1;
}
kmr_free_kvs(kvs0);
delete_file(common0.rank, common0.iteration + 1);
print_time(itr_times, ITERATIONS, rank);
kmr_free_context(mr);
kmr_fin();
MPI_Finalize();
return 0;
}