int main(int argc, char **argv)
{
gasnet_node_t rank, size;
size_t segsz = GASNET_PAGESIZE;
size_t heapsz = GASNET_PAGESIZE;
int argi;
GASNET_SAFE(gasnet_init(&argc, &argv));
rank = gasnet_mynode();
size = gasnet_nodes();
argi = 1;
if ((argi < argc) && !strcmp(argv[argi], "-m")) {
segsz = gasnet_getMaxLocalSegmentSize();
++argi;
}
GASNET_SAFE(gasnet_attach(NULL, 0, segsz, heapsz));
/* Only first and last print here, to keep managable I/O volume at scale */
if (!rank || (rank == size-1))
printf("Hello from node %d of %d\n", (int)rank, (int)size);
/* Spec says client should include a barrier before gasnet_exit() */
gasnet_barrier_notify(0,GASNET_BARRIERFLAG_ANONYMOUS);
gasnet_barrier_wait(0,GASNET_BARRIERFLAG_ANONYMOUS);
gasnet_exit(0);
/* Not reached in most implementations */
return 0;
}
void grt_init(int argc, char **argv) {
pthread_t thread;
/* Set up handler table */
if (!entry_table) {
entry_table = grt_entry_table;
table_size = GRT_TABLE_SIZE;
}
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
/* get SPMD info */
grt_id = gasnet_mynode();
grt_num_procs = gasnet_nodes();
gethostname(grt_proc_name, MAX_PROCESSOR_NAME);
/* Attach to network */
GASNET_Safe(gasnet_attach(entry_table, table_size,
GASNET_HEAP_SIZE, MINHEAPOFFSET));
if (grt_id == 0) {
printf("%s\n", argv[0]);
#ifdef GRT_WORD_32
printf("We are on a 32-bit machine.\n");
#else
printf("We are on a 64-bit machine.\n");
#endif
printf("gasnet_AMMaxMedium()=%lu\n", gasnet_AMMaxMedium());
printf("gasnet_AMMaxLongRequest()=%lu\n", gasnet_AMMaxLongRequest());
printf("gasnet_AMMaxLongReply()=%lu\n", gasnet_AMMaxLongReply());
}
fflush(stdout);
BARRIER();
fflush(stdout);
BARRIER();
/* Get segment info */
grt_seginfo = (gasnet_seginfo_t*) malloc(sizeof(gasnet_seginfo_t) *
grt_num_procs);
GASNET_Safe(gasnet_getSegmentInfo(grt_seginfo, grt_num_procs));
/* Initialize the heap for memory allocation */
grt_heap_base = grt_addr(grt_id, 0);
grt_heap = umalloc_makeheap(grt_heap_base,
grt_heap_size, UMALLOC_HEAP_GROWS_UP);
/* Spawn off a thread to handle remote handler requests */
pthread_create(&thread, NULL, poll, NULL);
/* Set up thread list */
linkedlist_init(&thread_list, 0, 0);
BARRIER();
}
int
main(int argc, char **argv)
{
int iters = 0;
gasnet_handlerentry_t htable[] = {
{ 201, chksum_reqh },
{ 202, chksum_reph }
};
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(htable, sizeof(htable)/sizeof(gasnet_handlerentry_t), TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testcore1",0,"(iters)");
assert(CHKSUM_TOTAL <= gasnet_AMMaxMedium());
if (argc > 1) iters = atoi(argv[1]);
if (!iters) iters = 1000;
if (argc > 2) test_usage();
/* get SPMD info */
chksum_iters = iters;
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
/* Only allow even number for numprocs */
if (numprocs % 2 != 0) {
MSG("WARNING: This test requires an even number of nodes. Test skipped.\n");
gasnet_exit(0); /* exit 0 to prevent false negatives in test harnesses for smp-conduit */
}
peerproc = (myproc % 2) ? myproc-1 : myproc+1;
seginfo_table = (gasnet_seginfo_t *) test_malloc(sizeof(gasnet_seginfo_t) * numprocs);
printf("%d> starting monoseed_init(%d)\n", myproc, iters);
monoseed_init(iters);
printf("%d> starting chksums_test(%d)\n", myproc, iters);
chksum_test(iters);
gasnet_exit(0);
return(0);
}
/* This tester measures the performance of a number of miscellaneous GASNet functions
that don't involve actual communication, to assist in evaluating the overhead of
the GASNet layer itself
*/
int main(int argc, char **argv) {
gasnet_handlerentry_t htable[] = {
{ hidx_null_shorthandler, null_shorthandler },
{ hidx_justreply_shorthandler, justreply_shorthandler },
{ hidx_null_medhandler, null_medhandler },
{ hidx_justreply_medhandler, justreply_medhandler },
{ hidx_null_longhandler, null_longhandler },
{ hidx_justreply_longhandler, justreply_longhandler }
};
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(htable, sizeof(htable)/sizeof(gasnet_handlerentry_t),
TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testmisc",1,"(iters) (accuracy_digits) (test_sections)");
mynode = gasnet_mynode();
myseg = TEST_MYSEG();
if (argc > 1) iters = atoi(argv[1]);
if (!iters) iters = 100000;
if (argc > 2) accuracy = atoi(argv[2]);
if (!accuracy) accuracy = 3;
if (argc > 3) TEST_SECTION_PARSE(argv[3]);
if (argc > 4) test_usage();
if (mynode == 0) {
printf("Running misc performance test with %i iterations...\n",iters);
printf("%-50s Total time Avg. time\n"
"%-50s ---------- ---------\n", "", "");
fflush(stdout);
}
doit1();
MSG("done.");
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
/* parse arguments */
arg = 1;
while (argc > arg) {
if (!strcmp(argv[arg], "-in")) {
insegment = 1;
++arg;
} else if (!strcmp(argv[arg], "-out")) {
insegment = 0;
++arg;
} else if (!strcmp(argv[arg], "-f")) {
firstlastmode = 1;
++arg;
} else if (!strcmp(argv[arg], "-a")) {
fullduplexmode = 1;
++arg;
} else if (!strcmp(argv[arg], "-p")) {
do_puts = 1; numflavors++;
++arg;
} else if (!strcmp(argv[arg], "-g")) {
do_gets = 1; numflavors++;
++arg;
} else if (!strcmp(argv[arg], "-s")) {
do_amshort = 1; numflavors++;
++arg;
} else if (!strcmp(argv[arg], "-m")) {
do_ammedium = 1; numflavors++;
++arg;
} else if (!strcmp(argv[arg], "-l")) {
do_amlong = 1; numflavors++;
++arg;
} else if (!strcmp(argv[arg], "-b")) {
do_bulk = 1;
++arg;
} else if (!strcmp(argv[arg], "-n")) {
do_nonbulk = 1;
++arg;
} else if (!strcmp(argv[arg], "-v")) {
do_value = 1;
++arg;
} else if (!strcmp(argv[arg], "-i")) {
do_implicit = 1; numsync++;
++arg;
} else if (!strcmp(argv[arg], "-e")) {
do_explicit = 1; numsync++;
++arg;
} else if (!strcmp(argv[arg], "-k")) {
do_blocking = 1; numsync++;
++arg;
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (fullduplexmode && firstlastmode) help = 1;
if (argc > arg+3) help = 1;
if (argc > arg) { iters = atoi(argv[arg]); arg++; }
if (!iters) iters = 10;
if (argc > arg) { maxdepth = atoi(argv[arg]); arg++; }
if (!maxdepth) maxdepth = 1024; /* 1024 default */
if (argc > arg) { maxsz = atoi(argv[arg]); arg++; }
if (!maxsz) maxsz = 2*1024*1024; /* 2 MB default */
#ifdef GASNET_SEGMENT_EVERYTHING
if (maxsz > TEST_SEGSZ) { MSG("maxsz must be <= %lu on GASNET_SEGMENT_EVERYTHING",(unsigned long)TEST_SEGSZ); gasnet_exit(1); }
#endif
GASNET_Safe(gasnet_attach(htable, sizeof(htable)/sizeof(gasnet_handlerentry_t),
TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testqueue",1,"[-in|-out|-a|-f] (iters) (maxdepth) (maxsz)\n"
" The 'in' or 'out' option selects whether the initiator-side\n"
" memory is in the GASNet segment or not (default is not).\n"
" The -a option enables full-duplex mode, where all nodes send.\n"
" The -f option enables 'first/last' mode, where the first node\n"
" sends to the last, while all other nodes sit idle.\n"
" Test types to run: (defaults to everything)\n"
" -p : puts\n"
" -g : gets\n"
" -s : AMShort\n"
" -m : AMMedium\n"
" -l : AMLong\n"
" -n : Test non-bulk put/gets\n"
" -b : Test bulk put/gets\n"
" -v : Test value-based put/gets\n"
" -i : Test implicit-handle put/gets\n"
" -e : Test explicit-handle put/gets\n"
" -k : Test blocking put/gets\n");
if (help) test_usage();
min_payload = 1;
max_payload = maxsz;
if (numflavors == 0) { /* default to all */
do_puts = 1;
do_gets = 1;
do_amshort = 1;
do_ammedium = 1;
do_amlong = 1;
}
if (numsync == 0) { /* default to all */
do_implicit = 1;
do_explicit = 1;
do_blocking = 1;
}
if (!do_bulk && !do_nonbulk && !do_value) {
do_bulk = 1;
do_nonbulk = 1;
do_value = 1;
}
if (!do_implicit && !do_explicit && !do_blocking) {
do_implicit = 1;
do_explicit = 1;
do_blocking = 1;
}
if (max_payload < min_payload) {
printf("ERROR: maxsz must be >= %i\n",min_payload);
gasnet_exit(1);
}
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
if (!firstlastmode) {
/* Only allow 1 or even number for numprocs */
if (numprocs > 1 && numprocs % 2 != 0) {
MSG0("WARNING: This test requires a unary or even number of nodes. Test skipped.\n");
gasnet_exit(0); /* exit 0 to prevent false negatives in test harnesses for smp-conduit */
}
}
/* Setting peer thread rank */
if (firstlastmode) {
peerproc = numprocs-1;
iamsender = (myproc == 0);
iamrecver = (myproc == numprocs-1);
multisender = 0;
} else if (numprocs == 1) {
peerproc = 0;
iamsender = 1;
iamrecver = 1;
multisender = 0;
} else {
peerproc = (myproc % 2) ? (myproc - 1) : (myproc + 1);
iamsender = (fullduplexmode || myproc % 2 == 0);
iamrecver = (fullduplexmode || !iamsender);
multisender = (fullduplexmode || numprocs >= 4);
}
multisender = 1; /* messes up output on some systems */
myseg = TEST_SEG(myproc);
tgtmem = TEST_SEG(peerproc);
if (insegment) {
msgbuf = (void *) myseg;
} else {
alloc = (void *) test_calloc(maxsz+PAGESZ,1); /* use calloc to prevent valgrind warnings */
msgbuf = (void *) alignup(((uintptr_t)alloc), PAGESZ); /* ensure page alignment of base */
}
assert(((uintptr_t)msgbuf) % PAGESZ == 0);
MSG0("Running %squeue test with local addr %sside segment, iters=%i, maxdepth=%i, sz: %i...%i",
firstlastmode ? "first/last " : (fullduplexmode ? "full-duplex ": ""),
insegment ? "in" : "out",
iters,
maxdepth,
min_payload, max_payload);
MSG0("x-axis: queue depth, y-axis: message size, injection time in microseconds\n");
BARRIER();
handles = (gasnet_handle_t *) test_malloc(sizeof(gasnet_handle_t) * maxdepth);
vghandles = (gasnet_valget_handle_t *) test_malloc(sizeof(gasnet_valget_handle_t) * maxdepth);
do_bulkputgets();
do_nonbulkputgets();
do_valueputgets();
do_blockingputgets();
do_amtests();
BARRIER();
test_free(handles);
if (!insegment) {
test_free(alloc);
}
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
int arg = 1, help = 0;
gasnet_handlerentry_t htable[] = {
{ hidx_ping_medhandler, ping_medhandler },
{ hidx_pong_medhandler, pong_medhandler },
{ hidx_ping_longhandler, ping_longhandler },
{ hidx_pong_longhandler, pong_longhandler },
{ hidx_ping_alonghandler, ping_alonghandler },
};
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
#define AMOPT() if (!amopt) { amopt = 1; domed = 0; dolong = 0; dolongasync = 0; }
while (argc > arg) {
if (!strcmp(argv[arg], "-p")) {
doprime = 1;
++arg;
} else if (!strcmp(argv[arg], "-u")) {
dosizesync = 0;
++arg;
} else if (!strcmp(argv[arg], "-s")) {
domultith = 0;
++arg;
} else if (!strcmp(argv[arg], "-n")) {
allowretry = 0;
++arg;
} else if (!strcmp(argv[arg], "-in")) {
doinseg = 1; dooutseg = 0;
++arg;
} else if (!strcmp(argv[arg], "-out")) {
doinseg = 0; dooutseg = 1;
++arg;
} else if (!strcmp(argv[arg], "-m")) {
AMOPT();
domed = 1;
++arg;
} else if (!strcmp(argv[arg], "-l")) {
AMOPT();
dolong = 1;
++arg;
} else if (!strcmp(argv[arg], "-a")) {
AMOPT();
dolongasync = 1;
++arg;
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (argc > arg) { iters = atoi(argv[arg]); arg++; }
if (!iters) iters = 10;
if (argc > arg) { max_payload = atoi(argv[arg]); arg++; }
if (!max_payload) max_payload = 1024*1024;
if (argc > arg) { depth = atoi(argv[arg]); arg++; }
if (!depth) depth = 16;
/* round down to largest payload AM allows */
maxlong = MIN(gasnet_AMMaxLongRequest(),gasnet_AMMaxLongReply());
max_payload = MIN(max_payload,MAX(gasnet_AMMaxMedium(),maxlong));
GASNET_Safe(gasnet_attach(htable, sizeof(htable)/sizeof(gasnet_handlerentry_t), TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testcore2",0,"[options] (iters) (max_payload) (depth)\n"
" -m test AMMedium (defaults to all types)\n"
" -l test AMLong (defaults to all types)\n"
" -a test AMLongAsync (defaults to all types)\n"
" -p prime the AMLong transfer areas with puts, to encourage pinning\n"
" -u loosen sychronization to allow diff payload sizes to be in flight at once\n"
" -s single-threaded PAR mode (default is to start a polling thread in PAR mode)\n"
" -n no retry on failure\n"
" -in/-out use only in- or out-of-segment sources for AMLong(Async) (default is both)\n"
);
if (help || argc > arg) test_usage();
TEST_PRINT_CONDUITINFO();
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
peerproc = myproc ^ 1;
if (peerproc == gasnet_nodes()) {
/* w/ odd # of nodes, last one talks to self */
peerproc = myproc;
}
myseg = TEST_MYSEG();
peerreqseg = TEST_SEG(peerproc);
peerrepseg = peerreqseg+max_payload*depth*2;
localseg = myseg + max_payload*depth*4;
assert_always(TEST_SEGSZ >= max_payload*depth*5);
privateseg = test_malloc(max_payload*depth*3); /* out-of-seg request src, long reply src, along reply src */
longreplysrc = privateseg+max_payload*depth;
alongreplysrc = privateseg+max_payload*depth*2;
#ifdef GASNET_PAR
if (domultith) test_createandjoin_pthreads(2,doit,NULL,0);
else
#endif
doit(0);
BARRIER();
test_free(privateseg);
MSG("done. (detected %i errs)", test_errs);
gasnet_exit(test_errs > 0 ? 1 : 0);
return 0;
}
int main(int argc, char **argv)
{
int iters = 0;
int arg;
void *myseg;
int firstlastmode = 0;
int fullduplexmode = 0;
int crossmachinemode = 0;
int skipwarmup = 0;
int help = 0;
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
/* parse arguments */
arg = 1;
while (argc > arg) {
if (!strcmp(argv[arg], "-f")) {
firstlastmode = 1;
++arg;
} else if (!strcmp(argv[arg], "-c")) {
crossmachinemode = 1;
++arg;
} else if (!strcmp(argv[arg], "-a")) {
fullduplexmode = 1;
++arg;
} else if (!strcmp(argv[arg], "-m")) {
unitsMB = 1;
++arg;
} else if (!strcmp(argv[arg], "-s")) {
skipwarmup = 1;
++arg;
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (argc > arg) { iters = atoi(argv[arg]); arg++; }
if (!iters) iters = 1000;
if (argc > arg) { maxsz = atoi(argv[arg]); arg++; }
if (!maxsz) maxsz = 1024*1024; /* 1 MB default */
if (argc > arg) { TEST_SECTION_PARSE(argv[arg]); arg++; }
#ifdef GASNET_SEGMENT_EVERYTHING
if (maxsz > TEST_SEGSZ) { ERR("maxsz must be <= %lu on GASNET_SEGMENT_EVERYTHING",(unsigned long)TEST_SEGSZ); gasnet_exit(1); }
#endif
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testmemset",1, "[options] (iters) (maxsz) (test_sections)\n"
" The -s option skips warm-up iterations\n"
" The -m option enables MB/sec units for bandwidth output (MB=2^20 bytes).\n"
" The -a option enables full-duplex mode, where all nodes send.\n"
" The -c option enables cross-machine pairing, default is nearest neighbor.\n"
" The -f option enables 'first/last' mode, where the first/last\n"
" nodes communicate with each other, while all other nodes sit idle.");
if (help || argc > arg) test_usage();
min_payload = 1;
max_payload = maxsz;
if (max_payload < min_payload) {
ERR("maxsz must be >= %i\n",min_payload);
test_usage();
}
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
if (!firstlastmode) {
/* Only allow 1 or even number for numprocs */
if (numprocs > 1 && numprocs % 2 != 0) {
MSG("WARNING: This test requires a unary or even number of nodes. Test skipped.\n");
gasnet_exit(0); /* exit 0 to prevent false negatives in test harnesses for smp-conduit */
}
}
/* Setting peer thread rank */
if (firstlastmode) {
peerproc = (myproc == 0 ? numprocs-1 : 0);
iamsender = (fullduplexmode ? myproc == 0 || myproc == numprocs-1 : myproc == 0);
} else if (numprocs == 1) {
peerproc = 0;
iamsender = 1;
} else if (crossmachinemode) {
if (myproc < numprocs / 2) {
peerproc = myproc + numprocs/2;
iamsender = 1;
} else {
peerproc = myproc - numprocs/2;
iamsender = fullduplexmode;
}
} else {
peerproc = (myproc % 2) ? (myproc - 1) : (myproc + 1);
iamsender = (fullduplexmode || myproc % 2 == 0);
}
myseg = TEST_SEG(myproc);
tgtmem = TEST_SEG(peerproc);
if (myproc == 0)
MSG("Running %i iterations of %s%s%smemset for sizes: %i...%i\n",
iters,
(firstlastmode ? "first/last " : ""),
(fullduplexmode ? "full-duplex ": ""),
(crossmachinemode ? "cross-machine ": ""),
min_payload, max_payload);
BARRIER();
if (iamsender && !skipwarmup) { /* pay some warm-up costs */
int i;
int warm_iters = MIN(iters, 32767); /* avoid hitting 65535-handle limit */
gasnet_handle_t *h = test_malloc(sizeof(gasnet_handle_t)*warm_iters);
for (i = 0; i < warm_iters; i++) {
gasnet_memset(peerproc, tgtmem, 0xff, 8);
gasnet_memset_nbi(peerproc, tgtmem, 0xff, 8);
h[i] = gasnet_memset_nb(peerproc, tgtmem, 0xff, 8);
}
gasnet_memset(peerproc, tgtmem, 0xff, max_payload);
gasnet_wait_syncnb_all(h, warm_iters);
gasnet_wait_syncnbi_puts();
test_free(h);
}
BARRIER();
if (TEST_SECTION_BEGIN_ENABLED()) bulk_test(iters);
if (TEST_SECTION_BEGIN_ENABLED()) bulk_test_nbi(iters);
if (TEST_SECTION_BEGIN_ENABLED()) bulk_test_nb(iters);
BARRIER();
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv)
{
int arg;
int iters = 0;
int size = 0;
int j;
int help = 0;
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testalign", 1,
"[-in|-out] (iters) (size)\n"
" The 'in' or 'out' option selects whether the initiator-side\n"
" memory is in the GASNet segment or not (default is not).\n"
" The -m option enables MB/sec units for bandwidth output (MB=2^20 bytes).");
/* parse arguments */
arg = 1;
while (argc > arg) {
if (!strcmp(argv[arg], "-in")) {
insegment = 1;
++arg;
} else if (!strcmp(argv[arg], "-out")) {
insegment = 0;
++arg;
} else if (!strcmp(argv[arg], "-m")) {
unitsMB = 1;
++arg;
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (help || argc > arg+2) test_usage();
if (argc > arg) iters = atoi(argv[arg++]);
if (!iters) iters = 1000;
if (argc > arg) size = atoi(argv[arg++]);
if (!size) size = DEFAULT_SZ;
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
/* Only allow even number for numprocs */
if (numprocs % 2 != 0) {
MSG("WARNING: This test requires an even number of nodes. Test skipped.\n");
gasnet_exit(0); /* exit 0 to prevent false negatives in test harnesses for smp-conduit */
}
/* Setting peer thread rank */
peerproc = (myproc % 2) ? (myproc - 1) : (myproc + 1);
iamsender = (myproc % 2 == 0);
rembuf = (void *) TEST_SEG(peerproc);
/* initialize global data in my thread */
if (insegment) {
locbuf = (void *)TEST_MYSEG();
} else {
/* size + 1 page of alignment + initial alignment padding of PAGESZ-1 */
uintptr_t tmp = (uintptr_t) test_malloc(size + 2 * PAGESZ - 1);
locbuf = (void *)((tmp + PAGESZ - 1) & ~(PAGESZ - 1));
}
for (j = 1; j <= PAGESZ; j *= 2) oneway_test(iters, size, j);
for (j = 1; j <= PAGESZ; j *= 2) oneway_nbi_test(iters, size, j);
for (j = 1; j <= PAGESZ; j *= 2) oneway_nb_test(iters, size, j);
BARRIER();
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv)
{
static int *A, *B, *C, *D, *E, *F, *G;
gasnet_node_t myproc, i;
int j;
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
if (argc > 1) {
iters = atoi(argv[1]);
}
if (iters < 1) {
iters = 1000;
}
#if GASNET_PAR
if (argc > 2) {
threads = atoi(argv[2]);
}
threads = test_thread_limit(threads);
if (threads < 1) {
printf("ERROR: Threads must be between 1 and %d\n", TEST_MAXTHREADS);
exit(EXIT_FAILURE);
}
#endif
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
images = numprocs * threads;
datasize = iters * (3 + 4 * images);
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testcoll",0,"(iters) (threadcnt)");
TEST_SET_WAITMODE(threads);
if (argc > 3) test_usage();
MSG0("Running coll test(s) with %d iterations.", iters);
R = test_malloc(iters*sizeof(int));
TEST_SRAND(1);
/* Number if ints to store */
/* Carve some variables out of the (aligned) segment: */
Aw = test_malloc(iters * sizeof(int *));
Bw = test_malloc(iters * sizeof(int *));
Cw = test_malloc(iters * sizeof(int *));
Dw = test_malloc(iters * sizeof(int *));
Ew = test_malloc(iters * sizeof(int *));
Fw = test_malloc(iters * sizeof(int *));
Gw = test_malloc(iters * sizeof(int *));
A = (int *)TEST_MYSEG(); /* int [1*iters] */
B = A + 1*iters; /* int [1*iters] */
C = B + 1*iters; /* int [N*iters] */
D = C + images*iters; /* int [N*iters] */
E = D + images*iters; /* int [1*iters] */
F = E + 1*iters; /* int [N*iters] */
G = F + images*iters; /* int [N*iters] */
for (j = 0; j < iters; ++j) {
Aw[j] = A + j;
Bw[j] = B + j;
Cw[j] = C + j*images;
Dw[j] = D + j*images;
Ew[j] = E + j;
Fw[j] = F + j*images;
Gw[j] = G + j*images;
}
/* The unaligned equivalents */
Av = test_malloc_2D(iters, images * sizeof(int *));
Bv = test_malloc_2D(iters, images * sizeof(int *));
Cv = test_malloc_2D(iters, images * sizeof(int *));
Dv = test_malloc_2D(iters, images * sizeof(int *));
Ev = test_malloc_2D(iters, images * sizeof(int *));
Fv = test_malloc_2D(iters, images * sizeof(int *));
Gv = test_malloc_2D(iters, images * sizeof(int *));
for (i = 0; i < images; ++i) {
/* Using (TEST_SEG(n) + n) yields unaligned even when the segments are aligned.
This is to help catch any case where addresses might have been misused that
might go undetected if the addresses were aligned */
A = (int *)TEST_SEG(i/threads) + (i/threads) + datasize*(i%threads);
B = A + 1*iters;
C = B + 1*iters;
D = C + images*iters;
E = D + images*iters;
F = E + 1*iters;
G = F + images*iters;
for (j = 0; j < iters; ++j) {
Av[j][i] = A + j;
Bv[j][i] = B + j;
Cv[j][i] = C + j*images;
Dv[j][i] = D + j*images;
Ev[j][i] = E + j;
Fv[j][i] = F + j*images;
Gv[j][i] = G + j*images;
}
}
BARRIER();
#if GASNET_PAR
MSG("Forking %d gasnet threads", threads);
{
int i;
thread_data_t* tt_thread_data = test_malloc(threads*sizeof(thread_data_t));
for (i = 0; i < threads; i++) {
tt_thread_data[i].myproc = myproc;
tt_thread_data[i].local_id = i;
tt_thread_data[i].mythread = i + threads * myproc;
tt_thread_data[i].peerthread = i + threads * (((myproc ^ 1) == numprocs) ? myproc : (myproc ^ 1));
}
test_createandjoin_pthreads(threads, &thread_main, tt_thread_data, sizeof(tt_thread_data[0]));
test_free(tt_thread_data);
}
#else
{ thread_data_t td;
td.myproc = myproc;
td.local_id = 0;
td.mythread = myproc;
td.peerthread = ((myproc ^ 1) == numprocs) ? myproc : (myproc ^ 1);
thread_main(&td);
}
#endif
BARRIER();
test_free(Aw);
test_free(Bw);
test_free(Cw);
test_free(Dw);
test_free_2D(Av);
test_free_2D(Bv);
test_free_2D(Cv);
test_free_2D(Dv);
test_free(R);
MSG("done.");
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv)
{
static int *A, *B, *C, *D, *E, *F, *G;
gasnet_node_t myproc, i;
int pollers = 0;
int j;
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
int arg = 1;
int help = 0;
while (argc > arg) {
if (!strcmp(argv[arg], "-p")) {
#if GASNET_PAR
++arg;
if (argc > arg) { pollers = atoi(argv[arg]); arg++; }
else help = 1;
#else
if (0 == gasnet_mynode()) {
fprintf(stderr, "testcoll %s\n", GASNET_CONFIG_STRING);
fprintf(stderr, "ERROR: The -p option is only available in the PAR configuration.\n");
fflush(NULL);
}
sleep(1);
gasnet_exit(1);
#endif
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (argc > arg) {
iters = atoi(argv[arg]);
++arg;
}
if (iters < 1) {
iters = 1000;
}
#if GASNET_PAR
if (argc > arg) {
threads = atoi(argv[arg]);
++arg;
}
threads = test_thread_limit(threads);
if (threads < 1) {
printf("ERROR: Threads must be between 1 and %d\n", TEST_MAXTHREADS);
exit(EXIT_FAILURE);
}
#endif
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
images = numprocs * threads;
datasize = iters * (3 + 4 * images);
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
#if GASNET_PAR
#define USAGE "[options] (iters) (threadcnt)\n" \
" The -p option gives the number of polling threads, specified as\n" \
" a non-negative integer argument (default is no polling threads).\n"
#else
#define USAGE "(iters)\n"
#endif
test_init("testcoll",0,USAGE);
TEST_SET_WAITMODE(threads + pollers);
if (argc > arg || help) test_usage();
MSG0("Running coll test(s) with %d iterations.", iters);
R = test_malloc(iters*sizeof(int));
/* Number if ints to store */
/* Carve some variables out of the (aligned) segment: */
Aw = test_malloc(iters * sizeof(int *));
Bw = test_malloc(iters * sizeof(int *));
Cw = test_malloc(iters * sizeof(int *));
Dw = test_malloc(iters * sizeof(int *));
Ew = test_malloc(iters * sizeof(int *));
Fw = test_malloc(iters * sizeof(int *));
Gw = test_malloc(iters * sizeof(int *));
A = (int *)TEST_MYSEG(); /* int [1*iters] */
B = A + 1*iters; /* int [1*iters] */
C = B + 1*iters; /* int [N*iters] */
D = C + images*iters; /* int [N*iters] */
E = D + images*iters; /* int [1*iters] */
F = E + 1*iters; /* int [N*iters] */
G = F + images*iters; /* int [N*iters] */
for (j = 0; j < iters; ++j) {
Aw[j] = A + j;
Bw[j] = B + j;
Cw[j] = C + j*images;
Dw[j] = D + j*images;
Ew[j] = E + j;
Fw[j] = F + j*images;
Gw[j] = G + j*images;
}
/* The unaligned equivalents */
Av = test_malloc_2D(iters, images * sizeof(int *));
Bv = test_malloc_2D(iters, images * sizeof(int *));
Cv = test_malloc_2D(iters, images * sizeof(int *));
Dv = test_malloc_2D(iters, images * sizeof(int *));
Ev = test_malloc_2D(iters, images * sizeof(int *));
Fv = test_malloc_2D(iters, images * sizeof(int *));
Gv = test_malloc_2D(iters, images * sizeof(int *));
for (i = 0; i < images; ++i) {
/* Using (TEST_SEG(n) + n) yields unaligned even when the segments are aligned.
This is to help catch any case where addresses might have been misused that
might go undetected if the addresses were aligned */
A = (int *)TEST_SEG(i/threads) + (i/threads) + datasize*(i%threads);
B = A + 1*iters;
C = B + 1*iters;
D = C + images*iters;
E = D + images*iters;
F = E + 1*iters;
G = F + images*iters;
for (j = 0; j < iters; ++j) {
Av[j][i] = A + j;
Bv[j][i] = B + j;
Cv[j][i] = C + j*images;
Dv[j][i] = D + j*images;
Ev[j][i] = E + j;
Fv[j][i] = F + j*images;
Gv[j][i] = G + j*images;
}
}
BARRIER();
#if GASNET_PAR
MSG("Forking %d gasnet threads (%d active, %d polling)", threads+pollers, threads, pollers);
{
int i;
thread_data_t* tt_thread_data = test_malloc(threads*sizeof(thread_data_t));
for (i = 0; i < threads; i++) {
tt_thread_data[i].myproc = myproc;
tt_thread_data[i].local_id = i;
tt_thread_data[i].mythread = i + threads * myproc;
tt_thread_data[i].peerthread = i + threads * (((myproc ^ 1) == numprocs) ? myproc : (myproc ^ 1));
}
test_createandjoin_pthreads(threads, &thread_main, tt_thread_data, sizeof(tt_thread_data[0]));
test_free(tt_thread_data);
}
#else
{ thread_data_t td;
td.myproc = myproc;
td.local_id = 0;
td.mythread = myproc;
td.peerthread = ((myproc ^ 1) == numprocs) ? myproc : (myproc ^ 1);
thread_main(&td);
}
#endif
BARRIER();
test_free(Aw);
test_free(Bw);
test_free(Cw);
test_free(Dw);
test_free_2D(Av);
test_free_2D(Bv);
test_free_2D(Cv);
test_free_2D(Dv);
test_free(R);
MSG("done.");
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
int i,j;
static uint8_t *A, *B;
thread_data_t *td_arr;
GASNET_Safe(gasnet_init(&argc, &argv));
performance_iters = DEFAULT_PERFORMANCE_ITERS;
for(i=1; i<argc; i++) {
if(strcmp("-i", argv[i])==0 || strcmp("-iters", argv[i])==0) {
performance_iters = atoi(argv[i+1]);
i++;
} else if(strcmp("-f", argv[i])==0 || strcmp("-tune-file", argv[i])==0) {
outputfile = test_malloc(strlen(argv[i+1])+1);
strcpy(outputfile, argv[i+1]);
i++;
} else if(strcmp("-p", argv[i])==0 || strcmp("-profile", argv[i])==0) {
profile_file = test_malloc(strlen(argv[i+1])+1);
strcpy(profile_file, argv[i+1]);
i++;
} else if(strcmp("-h", argv[i])==0 || strcmp("-help", argv[i])==0) {
MSG0("usage: %s (-i iters) (-f output file)\n", argv[0]);
gasnet_exit(0);
}
}
max_data_size = 0;
threads_per_node = 0;
profile_info = load_file(profile_file);
fill_meta_data(profile_info);
printf("data loaded: %d %d\n", (int)max_data_size, threads_per_node);
if(max_data_size <=0) {
fprintf(stderr, "max data size(%d) should be >=0\n", (int) max_data_size);
gasnet_exit(1);
}
if(threads_per_node<1) {
fprintf(stderr, "threads_per_node (%d) should be > 0\n", threads_per_node);
gasnet_exit(1);
}
if(performance_iters <=0) {
gasnet_exit(0);
}
mynode = gasnet_mynode();
nodes = gasnet_nodes();
THREADS = nodes * threads_per_node;
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
/* ?? test_init("testcolltuner",0,"(-i iters) (-f output_file)"); */
A = TEST_MYSEG();
B = A+(SEG_PER_THREAD*threads_per_node);
my_srcs = (uint8_t**) test_malloc(sizeof(uint8_t*)*threads_per_node);
my_dsts = (uint8_t**) test_malloc(sizeof(uint8_t*)*threads_per_node);
all_srcs = (uint8_t**) test_malloc(sizeof(uint8_t*)*THREADS);
all_dsts = (uint8_t**) test_malloc(sizeof(uint8_t*)*THREADS);
td_arr = (thread_data_t*) test_malloc(sizeof(thread_data_t)*threads_per_node);
TEST_SET_WAITMODE(threads_per_node);
for(i=0; i<threads_per_node; i++) {
my_srcs[i] = A + i*SEG_PER_THREAD;
my_dsts[i] = B + i*SEG_PER_THREAD;
td_arr[i].my_local_thread = i;
td_arr[i].mythread = mynode*threads_per_node+i;
td_arr[i].mysrc = my_srcs[i];
td_arr[i].mydest = my_dsts[i];
}
for(i=0; i<nodes; i++) {
/* assert_always(TEST_SEG(i).size >= SEG_PER_THREAD*threads_per_node); */
for(j=0; j<threads_per_node; j++) {
all_srcs[i*threads_per_node+j] = (uint8_t*) TEST_SEG(i) + j*SEG_PER_THREAD;
all_dsts[i*threads_per_node+j] = (uint8_t*) TEST_SEG(i) + SEG_PER_THREAD*threads_per_node + j*SEG_PER_THREAD;
}
}
#if GASNET_PAR
test_createandjoin_pthreads(threads_per_node, &thread_main, td_arr, sizeof(thread_data_t));
#else
thread_main(&td_arr[0]);
#endif
test_free(td_arr);
BARRIER();
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
int iters = 0;
int arg;
void *alloc = NULL;
int firstlastmode = 0;
int fullduplexmode = 0;
int crossmachinemode = 0;
int singlesender = 0;
int help = 0;
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
/* parse arguments */
arg = 1;
while (argc > arg) {
if (!strcmp(argv[arg], "-in")) {
insegment = 1;
++arg;
} else if (!strcmp(argv[arg], "-out")) {
insegment = 0;
++arg;
} else if (!strcmp(argv[arg], "-sl")) {
++arg;
if (argc > arg) { stridelevels = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-mincontig")) {
++arg;
if (argc > arg) { min_contig = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-maxcontig")) {
++arg;
if (argc > arg) { max_contig = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-contigfactor")) {
++arg;
if (argc > arg) { contigfactor = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-mindata")) {
++arg;
if (argc > arg) { min_payload = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-maxdata")) {
++arg;
if (argc > arg) { max_payload = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-datafactor")) {
++arg;
if (argc > arg) { datafactor = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-densitysteps")) {
++arg;
if (argc > arg) { densitysteps = atoi(argv[arg]); arg++; }
else help = 1;
} else if (!strcmp(argv[arg], "-f")) {
firstlastmode = 1;
++arg;
} else if (!strcmp(argv[arg], "-c")) {
crossmachinemode = 1;
++arg;
} else if (!strcmp(argv[arg], "-a")) {
fullduplexmode = 1;
++arg;
} else if (!strcmp(argv[arg], "-p")) {
dogets = 0; doputs = 1;
++arg;
} else if (!strcmp(argv[arg], "-g")) {
dogets = 1; doputs = 0;
++arg;
} else if (!strcmp(argv[arg], "-r")) {
remotecontig = 1;
++arg;
} else if (!strcmp(argv[arg], "-l")) {
localcontig = 1;
++arg;
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (argc > arg) { iters = atoi(argv[arg]); arg++; }
if (!iters) iters = 1000;
if (argc > arg) { TEST_SECTION_PARSE(argv[arg]); arg++; }
if (min_contig && max_contig && min_contig > max_contig) { ERR("min_contig > max_contig"); help = 1; }
if (min_payload && max_payload && min_payload > max_payload) { ERR("min_payload > max_payload"); help = 1; }
if (min_payload && min_contig && min_payload < min_contig) { ERR("min_payload < min_contig"); help = 1; }
if (max_contig && max_payload && max_contig > max_payload) { ERR("max_contig > max_payload"); help = 1; }
if (contigfactor < 2) { ERR("contigfactor < 2"); help = 1; }
if (datafactor < 2) { ERR("datafactor < 2"); help = 1; }
if (!max_payload) max_payload = 2*1024*1024; /* 2 MB default */
#ifdef GASNET_SEGMENT_EVERYTHING
maxsz = gasnet_getMaxGlobalSegmentSize();
#else
maxsz = 16*1024*1024;
#endif
max_payload = (int)MIN(maxsz, max_payload);
maxsz = MIN(((uint64_t)max_payload) * densitysteps,maxsz);
if (!min_contig) min_contig = 8;
if (!max_contig) max_contig = MIN(256*1024,max_payload);
if (!min_payload) min_payload = min_contig;
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testvisperf",1, "[options] (iters) (test_sections)\n"
" -p/-g selects puts only or gets only (default is both).\n"
" -r/-l selects remotely contiguous or locally contiguous (default is neither).\n"
" -mindata/-maxdata <sz> \n"
" selects sz as min/max data payload per operation.\n"
" -mincontig/-maxcontig <sz> \n"
" selects sz as min/max contig size.\n"
" -datafactor/-contigfactor <f> \n"
" selects f as growth factor for data/contig sizes.\n"
" -densitysteps <d> \n"
" selects d density steps, inclusive from 100%..100/d%\n"
" -sl <n> selects n striding levels (default is 2).\n"
" -in/-out selects whether the initiator-side\n"
" memory is in the GASNet segment or not (default is not).\n"
" -a enables full-duplex mode, where all nodes send.\n"
" -c enables cross-machine pairing, default is nearest neighbor.\n"
" -f enables 'first/last' mode, where the first/last\n"
" nodes communicate with each other, while all other nodes sit idle.");
if (help || argc > arg) test_usage();
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
if (!firstlastmode) {
/* Only allow 1 or even number for numprocs */
if (numprocs > 1 && numprocs % 2 != 0) {
MSG0("WARNING: This test requires a unary or even number of nodes. Test skipped.\n");
gasnet_exit(0); /* exit 0 to prevent false negatives in test harnesses for smp-conduit */
}
}
/* Setting peer thread rank */
if (firstlastmode) {
peerproc = (myproc == 0 ? numprocs-1 : 0);
iamsender = (fullduplexmode ? myproc == 0 || myproc == numprocs-1 : myproc == 0);
} else if (numprocs == 1) {
peerproc = 0;
iamsender = 1;
} else if (crossmachinemode) {
if (myproc < numprocs / 2) {
peerproc = myproc + numprocs/2;
iamsender = 1;
} else {
peerproc = myproc - numprocs/2;
iamsender = fullduplexmode;
}
} else {
peerproc = (myproc % 2) ? (myproc - 1) : (myproc + 1);
iamsender = (fullduplexmode || myproc % 2 == 0);
}
singlesender = (numprocs == 1) || ((numprocs == 2 || firstlastmode) && !fullduplexmode);
Rbase = TEST_SEG(peerproc);
if (insegment) {
Lbase = TEST_SEG(myproc);
} else {
alloc = test_calloc(maxsz+PAGESZ,1); /* use calloc to prevent valgrind warnings */
Lbase = alignup_ptr(alloc, PAGESZ); /* ensure page alignment of base */
}
assert(((uintptr_t)Lbase) % PAGESZ == 0);
if (myproc == 0) {
MSG0("Running %i iterations of %s%s%snon-contiguous put/get%s%s\n local data %s-segment for sizes: %i...%i\n",
iters,
(firstlastmode ? "first/last " : ""),
(fullduplexmode ? "full-duplex ": ""),
(crossmachinemode ? "cross-machine ": ""),
(remotecontig?"(remotely-contiguous)":""),
(localcontig?"(locally-contiguous)":""),
insegment ? "in" : "out",
min_payload, max_payload);
printf("rows are databytes/op : bandwidth values in MB/s\n");
}
BARRIER();
{ int contigsz;
int rawdatasz;
int isget;
test_vis_t viscat;
for (viscat = TEST_V; viscat <= TEST_S; viscat++) {
for (isget = 0; isget < 2; isget++) {
if (TEST_SECTION_BEGIN_ENABLED()) {
if (isget && !dogets) continue;
if (!isget && !doputs) continue;
if (!dovis[viscat]) continue;
for (contigsz = min_contig; contigsz <= max_contig; contigsz *= contigfactor) {
int di;
size_t lastdatasz = 0;
if (contigsz > max_payload) continue;
if (!myproc) {
printf("\n%c: %s %s CONTIGSZ = %i\n", TEST_SECTION_NAME(),
visdesc[(int)viscat], (isget?"GET":"PUT"), contigsz);
printf(" density:");
for (di = 0; di < densitysteps; di++) {
printf("%8i%%", (int)((densitysteps-di)*100.0/densitysteps));
}
printf("\n");
}
for (rawdatasz = min_payload; rawdatasz <= max_payload; rawdatasz *= datafactor) {
char mystr[255];
size_t datasz = aligndown(rawdatasz,contigsz);
if (datasz == lastdatasz) continue;
lastdatasz = datasz;
if (singlesender) snprintf(mystr, sizeof(mystr), "%8i: ", (int)datasz);
else snprintf(mystr, sizeof(mystr), "P%i: %6i: ", myproc, (int)datasz);
for (di = 0; di < densitysteps; di++) {
gasnett_tick_t begin=0, end=0;
size_t Lcnt = (localcontig ? 1 : datasz/contigsz);
size_t Rcnt = (remotecontig? 1 : datasz/contigsz);
size_t Lsz = datasz/Lcnt;
size_t Rsz = datasz/Rcnt;
void **Lilist = NULL;
void **Rilist = NULL;
gasnet_memvec_t *Lvlist = NULL;
gasnet_memvec_t *Rvlist = NULL;
size_t *Lstrides = NULL;
size_t *Rstrides = NULL;
size_t *LRcount = NULL;
size_t stride = contigsz*(((double)densitysteps)/(densitysteps-di));
if (stride * MAX(Lcnt,Rcnt) > maxsz) { strcat(mystr," - "); continue; }
if (iamsender) { /* setup metadata */
switch (viscat) {
case TEST_V:
Lvlist = make_vlist(Lbase, stride, Lcnt, Lsz);
Rvlist = make_vlist(Rbase, stride, Rcnt, Rsz);
break;
case TEST_I:
Lilist = make_ilist(Lbase, stride, Lcnt, Lsz);
Rilist = make_ilist(Rbase, stride, Rcnt, Rsz);
break;
case TEST_S: {
size_t chunkcnt = datasz/contigsz;
int dim;
Lstrides = test_malloc(sizeof(size_t)*stridelevels);
Rstrides = test_malloc(sizeof(size_t)*stridelevels);
LRcount = test_malloc(sizeof(size_t)*(stridelevels+1));
LRcount[0] = contigsz;
Lstrides[0] = (localcontig ? contigsz : stride);
Rstrides[0] = (remotecontig ? contigsz : stride);
for (dim = 1; dim < stridelevels; dim++) {
size_t factor = 1, fi;
for (fi = 1; fi <= chunkcnt/(2*(stridelevels-dim)); fi++) /* choose a reasonable factor */
if (chunkcnt/fi*fi == chunkcnt) factor = fi;
LRcount[dim] = factor;
chunkcnt /= factor;
Lstrides[dim] = LRcount[dim]*Lstrides[dim-1];
Rstrides[dim] = LRcount[dim]*Rstrides[dim-1];
}
LRcount[stridelevels] = chunkcnt;
{ size_t tmp = 1;
for (dim = 0; dim <= stridelevels; dim++) tmp *= LRcount[dim];
assert(tmp == datasz);
}
break;
}
}
}
#define DOIT(iters) do { \
int i; \
switch (viscat) { \
case TEST_V: \
for (i = 0; i < iters; i++) { \
if (isget) gasnet_getv_nbi_bulk(Lcnt,Lvlist,peerproc,Rcnt,Rvlist); \
else gasnet_putv_nbi_bulk(peerproc,Rcnt,Rvlist,Lcnt,Lvlist); \
} \
break; \
case TEST_I: \
for (i = 0; i < iters; i++) { \
if (isget) gasnet_geti_nbi_bulk(Lcnt,Lilist,Lsz,peerproc,Rcnt,Rilist,Rsz); \
else gasnet_puti_nbi_bulk(peerproc,Rcnt,Rilist,Rsz,Lcnt,Lilist,Lsz); \
} \
break; \
case TEST_S: \
for (i = 0; i < iters; i++) { \
if (isget) gasnet_gets_nbi_bulk(Lbase,Lstrides,peerproc,Rbase,Rstrides, \
LRcount,stridelevels); \
else gasnet_puts_nbi_bulk(peerproc,Rbase,Rstrides,Lbase,Lstrides, \
LRcount,stridelevels); \
} \
break; \
} \
gasnet_wait_syncnbi_all(); \
} while (0)
if (iamsender) DOIT(1); /* pay some warm-up costs */
BARRIER();
if (iamsender) {
begin = gasnett_ticks_now();
DOIT(iters);
end = gasnett_ticks_now();
}
BARRIER();
if (iamsender) {
char tmp[80];
double secs = gasnett_ticks_to_ns(end - begin)/1.0E9;
double dataMB = ((double)datasz) * iters / (1024*1024);
snprintf(tmp, sizeof(tmp), " %8.3f", dataMB / secs);
strcat(mystr, tmp);
}
if (Lilist) test_free(Lilist);
if (Rilist) test_free(Rilist);
if (Lvlist) test_free(Lvlist);
if (Rvlist) test_free(Rvlist);
if (Lstrides) test_free(Lstrides);
if (Rstrides) test_free(Rstrides);
if (LRcount) test_free(LRcount);
}
if (iamsender) { printf("%s\n", mystr); fflush(stdout); }
BARRIER();
}
}
}
}
}
}
BARRIER();
if (alloc) test_free(alloc);
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
int i = 0;
thread_data_t *td_arr;
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
mynode = gasnet_mynode();
nodes = gasnet_nodes();
if (argc > 1) iters = atoi(argv[1]);
if (!iters) iters = 10000;
#if GASNET_PAR
test_init("testteambarrier", 2, "(iters) (threadcount) (test sections)");
if (argc > 2) {
threads_per_node = atoi(argv[2]);
} else {
if (gasnett_getenv_yesno_withdefault("GASNET_TEST_POLITE_SYNC",0)) {
/* May overcommit only if somebody already expected it */
threads_per_node = gasnett_cpu_count();
} else {
threads_per_node = gasnett_cpu_count() / TEST_LOCALPROCS();
}
threads_per_node = MIN(threads_per_node, 8);
threads_per_node = test_thread_limit(threads_per_node);
threads_per_node = MAX(threads_per_node, 1);
}
if (threads_per_node > TEST_MAXTHREADS || threads_per_node < 1) {
printf("ERROR: Threads must be between 1 and %d\n", TEST_MAXTHREADS);
exit(EXIT_FAILURE);
}
if (argc > 3) TEST_SECTION_PARSE(argv[3]);
if (argc > 4) test_usage();
#else
test_init("testteambarrier", 1, "(iters) (test sections)");
threads_per_node = 1;
if (argc > 3) test_usage();
if (argc > 2) TEST_SECTION_PARSE(argv[2]);
#endif
TEST_SET_WAITMODE(threads_per_node);
td_arr = (thread_data_t*) test_malloc(sizeof(thread_data_t)*threads_per_node);
for(i=0; i<threads_per_node; i++) {
td_arr[i].my_local_thread = i;
td_arr[i].mythread = mynode*threads_per_node+i;
}
#if GASNET_PAR
test_createandjoin_pthreads(threads_per_node, &thread_main, td_arr, sizeof(thread_data_t));
#else
thread_main(&td_arr[0]);
#endif
test_free(td_arr);
gasnet_coll_barrier_notify(GASNET_TEAM_ALL, 0, GASNET_BARRIERFLAG_ANONYMOUS);
GASNET_Safe(gasnet_coll_barrier_wait(GASNET_TEAM_ALL, 0, GASNET_BARRIERFLAG_ANONYMOUS));
MSG("done.");
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
int help=0;
int arg=1;
gasnet_handlerentry_t htable[] = {
{ hidx_ping_shorthandler, ping_shorthandler },
{ hidx_pong_shorthandler, pong_shorthandler },
{ hidx_ping_medhandler, ping_medhandler },
{ hidx_pong_medhandler, pong_medhandler },
{ hidx_ping_longhandler, ping_longhandler },
{ hidx_pong_longhandler, pong_longhandler },
{ hidx_ping_shorthandler_flood, ping_shorthandler_flood },
{ hidx_pong_shorthandler_flood, pong_shorthandler_flood },
{ hidx_ping_medhandler_flood, ping_medhandler_flood },
{ hidx_pong_medhandler_flood, pong_medhandler_flood },
{ hidx_ping_longhandler_flood, ping_longhandler_flood },
{ hidx_pong_longhandler_flood, pong_longhandler_flood },
{ hidx_done_shorthandler, done_shorthandler }
};
GASNET_Safe(gasnet_init(&argc, &argv));
mynode = gasnet_mynode();
numnode = gasnet_nodes();
arg = 1;
while (argc > arg) {
if (!strcmp(argv[arg], "-p")) {
#if GASNET_PAR
pollers = atoi(argv[arg+1]);
arg += 2;
#else
if (0 == mynode) {
fprintf(stderr, "testam %s\n", GASNET_CONFIG_STRING);
fprintf(stderr, "ERROR: The -p option is only available in the PAR configuration.\n");
fflush(NULL);
}
sleep(1);
gasnet_exit(1);
#endif
} else if (!strcmp(argv[arg], "-in")) {
insegment = 1;
++arg;
} else if (!strcmp(argv[arg], "-out")) {
insegment = 0;
++arg;
} else if (!strcmp(argv[arg], "-c")) {
crossmachinemode = 1;
++arg;
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (argc > arg) { iters = atoi(argv[arg]); ++arg; }
if (!iters) iters = 1000;
if (argc > arg) { maxsz = atoi(argv[arg]); ++arg; }
if (!maxsz) maxsz = 2*1024*1024;
if (argc > arg) { TEST_SECTION_PARSE(argv[arg]); ++arg; }
GASNET_Safe(gasnet_attach(htable, sizeof(htable)/sizeof(gasnet_handlerentry_t),
TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
#if GASNET_PAR
test_init("testam", 1, "[options] (iters) (maxsz) (test_sections)\n"
" The '-in' or '-out' option selects whether the requestor's\n"
" buffer is in the GASNet segment or not (default is 'in').\n"
" The -p option gives the number of polling threads, specified as\n"
" a non-negative integer argument (default is no polling threads).\n"
" The -c option enables cross-machine pairing (default is nearest neighbor).\n");
#else
test_init("testam", 1, "[options] (iters) (maxsz) (test_sections)\n"
" The '-in' or '-out' option selects whether the requestor's\n"
" buffer is in the GASNet segment or not (default is 'in').\n"
" The -c option enables cross-machine pairing (default is nearest neighbor).\n");
#endif
if (help || argc > arg) test_usage();
TEST_PRINT_CONDUITINFO();
if (insegment) {
myseg = TEST_MYSEG();
} else {
char *space = test_malloc(alignup(maxsz,PAGESZ) + PAGESZ);
myseg = alignup_ptr(space, PAGESZ);
}
maxmed = MIN(maxsz, gasnet_AMMaxMedium());
maxlongreq = MIN(maxsz, gasnet_AMMaxLongRequest());
maxlongrep = MIN(maxsz, gasnet_AMMaxLongReply());
if (crossmachinemode) {
if ((numnode%2) && (mynode == numnode-1)) {
sender = 1;
peer = mynode;
} else {
gasnet_node_t half = numnode / 2;
sender = (mynode < half);
peer = sender ? (mynode + half) : (mynode - half);
}
} else {
peer = mynode ^ 1;
sender = mynode % 2 == 0;
if (peer == numnode) {
peer = mynode;
}
}
recvr = !sender || (peer == mynode);
peerseg = TEST_SEG(peer);
BARRIER();
if (mynode == 0) {
printf("Running %sAM performance test with %i iterations"
#if GASNET_PAR
" and %i extra recvr polling threads"
#endif
"...\n",
(crossmachinemode ? "cross-machine ": ""),
iters
#if GASNET_PAR
,pollers
#endif
);
printf(" Msg Sz Description Total time Avg. time Bandwidth\n"
" ------ ----------- ---------- --------- ---------\n");
fflush(stdout);
}
#if GASNET_PAR
if (pollers)
test_createandjoin_pthreads(pollers+1,doAll,NULL,0);
else
#endif
doAll(NULL);
MSG("done.");
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
struct delay_s {
int64_t delay_us;
int64_t delay_loops;
} delay_params;
int mynode, nodes, iters=0;
int64_t start,total,delay_us,baseline_us;
int64_t min_time, max_time, avg_time;
int64_t delay_loops = 0;
int j, i = 0;
int pause_len;
int pollcnt = 0;
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
test_init("testbarrierlate",1,"(iters) (pollcnt)");
mynode = gasnet_mynode();
nodes = gasnet_nodes();
if (argc > 1) iters = atoi(argv[1]);
if (!iters) iters = 10000;
if (argc > 2) pollcnt = atoi(argv[2]);
if (argc > 3) test_usage();
if (mynode == 0) {
printf("Running barrier late arrival test with %i iterations, pollcnt=%i...\n",iters, pollcnt);
fflush(stdout);
}
BARRIER();
/* warmup */
for (i=0; i < MIN(100,iters/100); i++) {
gasnet_barrier_notify(0, GASNET_BARRIERFLAG_ANONYMOUS);
GASNET_Safe(gasnet_barrier_wait(0, GASNET_BARRIERFLAG_ANONYMOUS));
}
BARRIER();
start = TIME();
for (i=0; i < iters; i++) {
gasnet_barrier_notify(0, GASNET_BARRIERFLAG_ANONYMOUS);
GASNET_Safe(gasnet_barrier_wait(0, GASNET_BARRIERFLAG_ANONYMOUS));
}
baseline_us = TIME() - start;
BARRIER();
if (mynode == 0) {
printf("Total time: %8.3f sec Avg Anon. Barrier latency: %8.3f us\n",
((float)baseline_us)/1000000, ((float)baseline_us)/iters);
fflush(stdout);
}
/* Calibrate a delay loop. Given "iters" and "delay_us", we determine the argument
* we need when calling test_delay() iters times, to get a _total_ delay no less than
* delay_us. The value of delay_us is overwritten with the achieved delay.
* We calibrate the delay on exactly one node to avoid spoiling timings on
* overcommitted CPUs. Nodes not performing the calibration sleep for at least twice
* the time we are calibrating for. (No way to be sure this is enough, since
* calibration is iterative.)
*/
BARRIER();
pause_len = 1 + 4 * (baseline_us + 999999)/1000000;
if (mynode == 0) {
struct delay_s *p = (struct delay_s *)TEST_MYSEG();
start = TIME();
printf("Calibrating delay loop (expect at least a %d sec pause)...\n", pause_len);
fflush(stdout);
p->delay_us = 2*baseline_us; /* delay at least two full barrier times */
p->delay_loops = test_calibrate_delay(iters, pollcnt, &(p->delay_us));
} else {
sleep(pause_len);
}
BARRIER();
gasnet_get(&delay_params, 0, TEST_SEG(0), sizeof(struct delay_s));
delay_us = delay_params.delay_us;
delay_loops = delay_params.delay_loops;
if (mynode == 0) {
printf("Calibration complete (actual pause = %5.3f sec).\n", (float)((TIME()-start)/1000000.0));
printf("Ideal loop time = %8.3f sec.\n", (float)(delay_us)/1000000.0);
fflush(stdout);
}
/* Take turns being late to notify
* We insert a delay before the _notify() on one node.
* This simulates a load imbalance between barriers.
* The time reported is how much the barrier costs excluding the delay.
* This reported time will often be less than the full barrier because
* some progress was made by the other nodes.
*/
avg_time = 0;
max_time = 0;
min_time = (int64_t)1 << 62; /* good enough */
for (j=0; j < nodes; j++) {
BARRIER();
start = TIME();
for (i=0; i < iters; i++) {
if (j == mynode) {
test_delay(delay_loops, pollcnt);
}
gasnet_barrier_notify(0, GASNET_BARRIERFLAG_ANONYMOUS);
GASNET_Safe(gasnet_barrier_wait(0, GASNET_BARRIERFLAG_ANONYMOUS));
}
total = TIME() - start;
if (mynode == 0) {
printf("Total time: %8.3f sec Late-notify test on node %d\n", ((float)total)/1000000, j);
fflush(stdout);
}
total -= delay_us;
avg_time += total;
min_time = MIN(min_time, total);
max_time = MAX(max_time, total);
}
avg_time /= nodes;
if (mynode == 0) {
printf("Total difference: %8.3f sec Late notify() Anon. Barrier net latency, minimum: %8.3f us (%6.2f%%)\n", ((float)min_time)/1000000, ((float)min_time)/iters, ((float)min_time * 100.)/baseline_us);
printf("Total difference: %8.3f sec Late notify() Anon. Barrier net latency, maximum: %8.3f us (%6.2f%%)\n", ((float)max_time)/1000000, ((float)max_time)/iters, ((float)max_time * 100.)/baseline_us);
printf("Total difference: %8.3f sec Late notify() Anon. Barrier net latency, average: %8.3f us (%6.2f%%)\n", ((float)avg_time)/1000000, ((float)avg_time)/iters, ((float)avg_time * 100.)/baseline_us);
fflush(stdout);
}
/* Take turns being late to wait
* We insert a delay between the _notify() and _wait() on one node.
* This simulates a load imbalance between barrier notify and wait.
* The time reported is how much the barrier costs excluding the delay.
* This reported time will often be less than the full barrier because
* some progress was made by the other nodes.
*/
avg_time = 0;
max_time = 0;
min_time = (int64_t)1 << 62; /* good enough */
for (j=0; j < nodes; j++) {
BARRIER();
start = TIME();
for (i=0; i < iters; i++) {
gasnet_barrier_notify(0, GASNET_BARRIERFLAG_ANONYMOUS);
if (j == mynode) {
test_delay(delay_loops, pollcnt);
}
GASNET_Safe(gasnet_barrier_wait(0, GASNET_BARRIERFLAG_ANONYMOUS));
}
total = TIME() - start;
if (mynode == 0) {
printf("Total time: %8.3f sec Late-wait test on node %d\n", ((float)total)/1000000, j);
fflush(stdout);
}
total -= delay_us;
avg_time += total;
min_time = MIN(min_time, total);
max_time = MAX(max_time, total);
}
avg_time /= nodes;
if (mynode == 0) {
printf("Total difference: %8.3f sec Late wait() Anon. Barrier net latency, minimum: %8.3f us (%6.2f%%)\n", ((float)min_time)/1000000, ((float)min_time)/iters, ((float)min_time * 100.)/baseline_us);
printf("Total difference: %8.3f sec Late wait() Anon. Barrier net latency, maximum: %8.3f us (%6.2f%%)\n", ((float)max_time)/1000000, ((float)max_time)/iters, ((float)max_time * 100.)/baseline_us);
printf("Total difference: %8.3f sec Late wait() Anon. Barrier net latency, average: %8.3f us (%6.2f%%)\n", ((float)avg_time)/1000000, ((float)avg_time)/iters, ((float)avg_time * 100.)/baseline_us);
fflush(stdout);
}
BARRIER();
MSG("done.");
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv)
{
int mynode, nodes, iters=0;
int64_t start,total;
int i = 0;
gasnet_node_t nrows, ncols, my_row, my_col;
void *clientdata = NULL;
gasnet_team_handle_t my_row_team, my_col_team;
static uint8_t *A, *B, *C;
static int *B_int, *C_int;
int num_iters;
size_t sz;
gasnet_seginfo_t teamA_scratch;
gasnet_seginfo_t teamB_scratch;
gasnet_seginfo_t const * test_segs;
GASNET_Safe(gasnet_init(&argc, &argv));
#if !GASNET_SEQ
MSG0("WARNING: This test does not work for NON-SEQ builds yet.. skipping test\n");
gasnet_exit(0);
#endif
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
A = TEST_MYSEG();
B = A + SCRATCH_SIZE;
C = B + SCRATCH_SIZE;
B_int = (int*) B;
C_int = (int*) C;
gasnet_coll_init(NULL, 0, NULL, 0, 0);
test_init("testteambcast", 1, "(nrows) (ncols) (iters)");
mynode = gasnet_mynode();
nodes = gasnet_nodes();
test_segs = TEST_SEGINFO();
teamA_scratch.addr = test_segs[mynode].addr;
teamA_scratch.size = SCRATCH_SIZE/2;
teamB_scratch.addr = (uint8_t*)teamA_scratch.addr + SCRATCH_SIZE/2;
teamB_scratch.size = teamA_scratch.size;
if (argc > 4)
test_usage();
if (argc > 1) {
nrows = atoi(argv[1]);
} else {
nrows = 1 + !(nodes & 1); /* 1 if odd, 2 if even */
}
if (argc > 2) {
ncols = atoi(argv[2]);
} else {
ncols = nodes / nrows;
}
assert_always(nrows*ncols == nodes);
if (argc > 3) iters = atoi(argv[3]);
if (!iters) iters = 20;
if (mynode == 0) {
printf("Running team test with a %u-by-%u grid and %i iterations...\n",
nrows, ncols, iters);
fflush(stdout);
}
BARRIER();
my_row = mynode / ncols;
my_col = mynode % ncols;
my_row_team = gasnet_coll_team_split(GASNET_TEAM_ALL,
my_row,
my_col,
&teamA_scratch);
my_col_team = gasnet_coll_team_split(GASNET_TEAM_ALL,
my_col,
my_row,
&teamB_scratch);
if (my_col == 0) {
printf("row team %u: Running team barrier test with row teams...\n",
my_row);
fflush(stdout);
}
BARRIER();
start = TIME();
for (i=0; i < iters*10; i++) {
gasnet_coll_barrier_notify(my_row_team, 0, GASNET_BARRIERFLAG_UNNAMED);
gasnet_coll_barrier_wait(my_row_team, 0, GASNET_BARRIERFLAG_UNNAMED);
}
total = TIME() - start;
if (my_col == 0) {
printf("row team %u: total time: %8.3f sec, avg row team Barrier latency: %8.3f us\n",
my_row, ((float)total)/1000000, ((float)total)/(iters*10));
fflush(stdout);
}
if (my_row == 0) {
printf("col team %u: Running team barrier test with column teams...\n",
my_col);
fflush(stdout);
}
BARRIER();
start = TIME();
for (i=0; i < iters*10; i++) {
gasnet_coll_barrier_notify(my_col_team, 0, GASNET_BARRIERFLAG_UNNAMED);
gasnet_coll_barrier_wait(my_col_team, 0, GASNET_BARRIERFLAG_UNNAMED);
}
total = TIME() - start;
if (my_row == 0) {
printf("col team %u: total time: %8.3f sec Avg column team Barrier latency: %8.3f us\n",
my_col, ((float)total)/1000000, ((float)total)/(iters*10));
fflush(stdout);
}
BARRIER();
/*first do team all broadcast*/
for (sz = 1; sz<MAX_SIZE; sz=sz*2) {
int root = 0;
for(i=0; i<sz; i++) {
B_int[i] = mynode*sz+42+i;
C_int[i] = -1;
}
BARRIER();
gasnet_coll_broadcast(GASNET_TEAM_ALL, C_int, root, B_int, sz*sizeof(int),
GASNET_COLL_IN_ALLSYNC|GASNET_COLL_OUT_NOSYNC|GASNET_COLL_LOCAL);
BARRIER();
for(i=0; i<sz; i++) {
int expected = root*sz+42+i;
if(expected != C_int[i]) {
fprintf(stderr, "%d> %d %d (expecting %d)\n", mynode, i, C_int[i], expected);
gasnet_exit(1);
}
}
BARRIER();
/*time this*/
start = TIME();
for(i=0; i<iters; i++) {
gasnet_coll_broadcast(GASNET_TEAM_ALL, C_int, root, B_int, sz*sizeof(int),
GASNET_COLL_IN_ALLSYNC|GASNET_COLL_OUT_NOSYNC|GASNET_COLL_LOCAL);
}
total = TIME() - start;
if(mynode == 0){
printf("%d> %ld byte broadcast team all time: %8.3f usec\n",mynode,(long)sz*sizeof(int),
((float)total)/(iters));
fflush(stdout);
}
}
BARRIER();
/*next do row broadcasts*/
for (sz = 1; sz<MAX_SIZE; sz=sz*2) {
for(i=0; i<sz; i++) {
B_int[i] = mynode*sz+42+i;
C_int[i] = -1;
}
BARRIER();
gasnet_coll_broadcast(my_row_team, C_int, 0, B_int, sz*sizeof(int),
GASNET_COLL_IN_ALLSYNC|GASNET_COLL_OUT_NOSYNC|GASNET_COLL_LOCAL);
BARRIER();
for(i=0; i<sz; i++) {
int expected = my_row*ncols*sz+42+i;
if(expected != C_int[i]) {
fprintf(stderr, "%d> %d %d (expecting %d)\n", mynode, i, C_int[i], expected);
gasnet_exit(1);
}
}
BARRIER();
/*time this*/
start = TIME();
for(i=0; i<iters; i++) {
gasnet_coll_broadcast(my_row_team, C_int, 0, B_int, sz*sizeof(int),
GASNET_COLL_IN_ALLSYNC|GASNET_COLL_OUT_NOSYNC|GASNET_COLL_LOCAL);
}
total = TIME() - start;
if(my_col == 0){
printf("%d> %ld byte broadcast row team %u time: %8.3f usec\n",mynode,(long)sz*sizeof(int),
my_row, ((float)total)/(iters));
fflush(stdout);
}
}
BARRIER();
/*next do col broadcasts*/
for (sz = 1; sz<MAX_SIZE; sz=sz*2) {
for(i=0; i<sz; i++) {
B_int[i] = mynode*sz+42+i;
C_int[i] = -1;
}
BARRIER();
gasnet_coll_broadcast(my_col_team, C_int, 0, B_int, sz*sizeof(int),
GASNET_COLL_IN_ALLSYNC|GASNET_COLL_OUT_NOSYNC|GASNET_COLL_LOCAL);
BARRIER();
for(i=0; i<sz; i++) {
int expected = my_col*sz+42+i;
if(expected != C_int[i]) {
fprintf(stderr, "%d> %d %d (expecting %d)\n", mynode, i, C_int[i], expected);
gasnet_exit(1);
}
}
BARRIER();
/*time this*/
start = TIME();
for(i=0; i<iters; i++) {
gasnet_coll_broadcast(my_col_team, C_int, 0, B_int, sz*sizeof(int),
GASNET_COLL_IN_ALLSYNC|GASNET_COLL_OUT_NOSYNC|GASNET_COLL_LOCAL);
}
total = TIME() - start;
if(my_row == 0){
printf("%d> %ld byte broadcast col team %u time: %8.3f usec\n",mynode,(long)sz*sizeof(int),
my_col, ((float)total)/(iters));
fflush(stdout);
}
}
BARRIER();
MSG("done.");
gasnet_exit(0); /* for faster exit */
return 0;
}
int main(int argc, char **argv)
{
int outer_iterations = 0;
int inner_iterations = 0;
int seedoffset = 0;
int numprocs, myproc, peerproc;
int sender_p;
char *shadow_region_1, *shadow_region_2;
int i,j;
char *local_base, *target_base;
/* call startup */
GASNET_Safe(gasnet_init(&argc, &argv));
/* get SPMD info */
myproc = gasnet_mynode();
numprocs = gasnet_nodes();
if (argc > 1) segsize = atoi(argv[1]);
if (!segsize) segsize = 1024*1000;
if (argc > 2) outer_iterations = atoi(argv[2]);
if (!outer_iterations) outer_iterations = 10;
if (argc > 3) inner_iterations = atoi(argv[3]);
if (!inner_iterations) inner_iterations = 10;
if (argc > 4) seedoffset = atoi(argv[4]);
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ, TEST_MINHEAPOFFSET));
test_init("testslice",0, "(segsize) (iterations) (# of sizes per iteration) (seed)");
/* parse arguments */
if (argc > 5) test_usage();
if(numprocs & 1) {
MSG0("WARNING: This test requires an even number of nodes. Test skipped.\n");
gasnet_exit(0); /* exit 0 to prevent false negatives in test harnesses for smp-conduit */
}
sender_p = !(myproc & 1);
peerproc = myproc ^ 1;
if (seedoffset == 0) {
seedoffset = (((unsigned int)TIME()) & 0xFFFF);
TEST_BCAST(&seedoffset, 0, &seedoffset, sizeof(&seedoffset));
}
TEST_SRAND(myproc+seedoffset);
MSG0("Running with segment size = %d outer iterations=%d inner iterations=%d seed=%d",
segsize,outer_iterations, inner_iterations, seedoffset);
BARRIER();
/* Allocate two shadow regions the same size as the segment */
shadow_region_1 = (char *) test_malloc(segsize);
shadow_region_2 = (char *) test_malloc(segsize);
/* Fill up the shadow region with random data */
for(i=0; i < segsize; i++) {
shadow_region_1[i] = (char) TEST_RAND(0,255);
}
memset(shadow_region_2,0,segsize);
/* Big loop performing the following */
for(i=0; i < outer_iterations; i++) {
if(sender_p) {
/* Pick a starting point anywhere in the segment */
int starting_point = TEST_RAND(0,(segsize-1));
local_base = TEST_SEG(myproc);
target_base = TEST_SEG(peerproc);
for(j=0; j < inner_iterations; j++) {
/* Pick a length */
int len = TEST_RAND(1,segsize-starting_point);
int remote_starting_point = TEST_RAND(0,segsize-len);
int local_starting_point_1 = TEST_RAND(0,segsize-len);
int local_starting_point_2 = TEST_RAND(0,segsize-len);
/* Perform operations */
/* Out of segment put from shadow_region 1 to remote */
gasnet_put(peerproc,target_base+remote_starting_point,shadow_region_1 + starting_point,len);
/* In segment get from remote to local segment */
gasnet_get(local_base+local_starting_point_1,peerproc,target_base+remote_starting_point,len);
/* Verify */
assert_eq(shadow_region_1 + starting_point, local_base + local_starting_point_1, len,starting_point,i,j,"Out of segment put + in segment get");
/* Out of segment get from remote to shadow_region_2 (starting from 0) */
gasnet_get(shadow_region_2+local_starting_point_2,peerproc,target_base+remote_starting_point,len);
/* Verify */
assert_eq(shadow_region_2+local_starting_point_2, shadow_region_1 + starting_point, len,starting_point,i,j,"Out of segment get");
}
TEST_PROGRESS_BAR(i,outer_iterations);
}
BARRIER();
}
if(sender_p && !failures) {
MSG("testslice PASSED");
}
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv) {
int help=0;
int arg=1;
gasnet_handlerentry_t htable[] = {
{ hidx_ping_shorthandler, ping_shorthandler },
{ hidx_pong_shorthandler, pong_shorthandler },
{ hidx_ping_medhandler, ping_medhandler },
{ hidx_pong_medhandler, pong_medhandler },
{ hidx_ping_longhandler, ping_longhandler },
{ hidx_pong_longhandler, pong_longhandler },
{ hidx_ping_shorthandler_flood, ping_shorthandler_flood },
{ hidx_pong_shorthandler_flood, pong_shorthandler_flood },
{ hidx_ping_medhandler_flood, ping_medhandler_flood },
{ hidx_pong_medhandler_flood, pong_medhandler_flood },
{ hidx_ping_longhandler_flood, ping_longhandler_flood },
{ hidx_pong_longhandler_flood, pong_longhandler_flood },
{ hidx_done_shorthandler, done_shorthandler }
};
GASNET_Safe(gasnet_init(&argc, &argv));
mynode = gasnet_mynode();
numnode = gasnet_nodes();
arg = 1;
while (argc > arg) {
if (!strcmp(argv[arg], "-p")) {
#if GASNET_PAR
pollers = test_thread_limit(atoi(argv[arg+1])+1)-1;
arg += 2;
#else
if (0 == mynode) {
fprintf(stderr, "testam %s\n", GASNET_CONFIG_STRING);
fprintf(stderr, "ERROR: The -p option is only available in the PAR configuration.\n");
fflush(NULL);
}
sleep(1);
gasnet_exit(1);
#endif
} else if (!strcmp(argv[arg], "-in")) {
insegment = 1;
++arg;
} else if (!strcmp(argv[arg], "-out")) {
insegment = 0;
++arg;
} else if (!strcmp(argv[arg], "-c")) {
crossmachinemode = 1;
++arg;
} else if (!strcmp(argv[arg], "-src-noop")) {
src_mode = SRC_NOOP;
++arg;
} else if (!strcmp(argv[arg], "-src-generate")) {
src_mode = SRC_GENERATE;
++arg;
} else if (!strcmp(argv[arg], "-src-memcpy")) {
src_mode = SRC_MEMCPY;
++arg;
} else if (argv[arg][0] == '-') {
help = 1;
++arg;
} else break;
}
if (argc > arg) { iters = atoi(argv[arg]); ++arg; }
if (!iters) iters = 1000;
if (argc > arg) { maxsz = atoi(argv[arg]); ++arg; }
if (!maxsz) maxsz = 2*1024*1024;
if (argc > arg) { TEST_SECTION_PARSE(argv[arg]); ++arg; }
GASNET_Safe(gasnet_attach(htable, sizeof(htable)/sizeof(gasnet_handlerentry_t),
TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
#if GASNET_PAR
#define PAR_USAGE \
" The -p option gives the number of polling threads, specified as\n" \
" a non-negative integer argument (default is no polling threads).\n"
#else
#define PAR_USAGE ""
#endif
test_init("testam", 1, "[options] (iters) (maxsz) (test_sections)\n"
" The '-in' or '-out' option selects whether the requestor's\n"
" buffer is in the GASNet segment or not (default is 'in').\n"
PAR_USAGE
" The '-src-*' options select treatment of the payload buffer used for\n"
" Medium and Long AMs, as follows:\n"
" -src-noop: no per-operation initialization (default)\n"
" -src-generate: initialized (w/o memory reads) on each AM injection\n"
" -src-memcpy: initialized using memcpy() on each AM injection\n"
" The -c option enables cross-machine pairing (default is nearest neighbor).\n");
if (help || argc > arg) test_usage();
TEST_PRINT_CONDUITINFO();
if (insegment) {
myseg = TEST_MYSEG();
} else {
char *space = test_malloc(alignup(maxsz,PAGESZ) + PAGESZ);
myseg = alignup_ptr(space, PAGESZ);
}
maxmed = MIN(maxsz, gasnet_AMMaxMedium());
maxlongreq = MIN(maxsz, gasnet_AMMaxLongRequest());
maxlongrep = MIN(maxsz, gasnet_AMMaxLongReply());
if (src_mode == SRC_MEMCPY) {
zero_buffer = test_calloc(maxsz, 1);
}
if (crossmachinemode) {
if ((numnode%2) && (mynode == numnode-1)) {
sender = 1;
peer = mynode;
} else {
gasnet_node_t half = numnode / 2;
sender = (mynode < half);
peer = sender ? (mynode + half) : (mynode - half);
}
} else {
peer = mynode ^ 1;
sender = mynode % 2 == 0;
if (peer == numnode) {
peer = mynode;
}
}
recvr = !sender || (peer == mynode);
// Long Request and Reply (distinct for loopback)
reply_addr = TEST_SEG(peer);
request_addr = (peer == mynode) ? (void*)((uintptr_t)reply_addr + alignup(maxsz,SIZEOF_GASNET_REGISTER_VALUE_T))
: reply_addr;
BARRIER();
#if GASNET_PAR
#define PAR_FMT " %i extra recvr polling threads\n"
#define PAR_ARG ,pollers
#else
#define PAR_FMT /*empty*/
#define PAR_ARG /*empty*/
#endif
if (mynode == 0) {
printf("Running %i iterations of %s AM performance with:\n"
" local addresses %sside the segment%s\n"
" %s\n"
PAR_FMT
" ...\n",
iters,
(crossmachinemode ? "cross-machine ": ""),
(insegment ? "in" : "out"),
(insegment ? " (default)" : ""),
((src_mode == SRC_NOOP) ? "no payload initialization (default)"
:(src_mode == SRC_GENERATE) ? "payload initialized by computation"
: "payload initialized using memcpy()")
PAR_ARG
);
printf(" Msg Sz Description Total time Avg. time Bandwidth\n"
" ------ ----------- ---------- --------- ---------\n");
fflush(stdout);
}
#if GASNET_PAR
TEST_SET_WAITMODE(pollers+1);
if (pollers)
test_createandjoin_pthreads(pollers+1,doAll,NULL,0);
else
#endif
doAll(NULL);
MSG("done.");
gasnet_exit(0);
return 0;
}
int main(int argc, char **argv)
{
int mynode, nodes, iters=0;
int64_t start,total;
int i = 0;
gasnet_node_t nrows, ncols, my_row, my_col;
void *clientdata = NULL;
gasnet_team_handle_t my_row_team, my_col_team;
static uint8_t *A, *B;
gasnet_seginfo_t teamA_scratch;
gasnet_seginfo_t teamB_scratch;
gasnet_seginfo_t const * test_segs;
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(NULL, 0, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
#if !GASNET_SEQ
MSG0("WARNING: This test does not work for NON-SEQ builds yet.. skipping test\n");
gasnet_exit(0);
#endif
A = TEST_MYSEG();
gasnet_coll_init(NULL, 0, NULL, 0, 0);
test_init("testteam", 1, "(iters) (nrows) (ncols)");
mynode = gasnet_mynode();
nodes = gasnet_nodes();
test_segs = TEST_SEGINFO();
teamA_scratch.addr = test_segs[mynode].addr;
teamA_scratch.size = test_segs[mynode].size/2;
teamB_scratch.addr = (uint8_t*)teamA_scratch.addr + teamA_scratch.size;
teamB_scratch.size = teamA_scratch.size;
if (argc > 4)
test_usage();
if (argc > 1) iters = atoi(argv[1]);
if (!iters) iters = 10000;
if (argc > 2) {
nrows = atoi(argv[2]);
} else {
/* search for as near to square as possible */
nrows = sqrt(nodes);
while (nodes % nrows) --nrows;
}
if (argc > 3) {
ncols = atoi(argv[3]);
} else {
ncols = nodes / nrows;
}
assert_always(nrows*ncols == nodes);
MSG0("Running team test with a %u-by-%u grid and %i iterations...\n",
(int)nrows, (int)ncols, iters);
BARRIER();
my_row = mynode / ncols;
my_col = mynode % ncols;
my_row_team = gasnet_coll_team_split(GASNET_TEAM_ALL,
my_row,
my_col,
&teamA_scratch);
my_col_team = gasnet_coll_team_split(GASNET_TEAM_ALL,
my_col,
my_row,
&teamB_scratch);
if (my_col == 0) {
printf("row team %u: Running team barrier test with row teams...\n",
(int)my_row);
fflush(stdout);
}
BARRIER();
start = TIME();
for (i=0; i < iters; i++) {
gasnete_coll_teambarrier_notify(my_row_team);
gasnete_coll_teambarrier_wait(my_row_team);
}
total = TIME() - start;
if (my_col == 0) {
printf("row team %u: total time: %8.3f sec, avg row team Barrier latency: %8.3f us\n",
(int)my_row, ((float)total)/1000000, ((float)total)/iters);
fflush(stdout);
}
if (my_row == 0) {
printf("col team %u: Running team barrier test with column teams...\n",
(int)my_col);
fflush(stdout);
}
BARRIER();
start = TIME();
for (i=0; i < iters; i++) {
gasnete_coll_teambarrier_notify(my_col_team);
gasnete_coll_teambarrier_wait(my_col_team);
}
total = TIME() - start;
if (my_row == 0) {
printf("col team %u: total time: %8.3f sec Avg column team Barrier latency: %8.3f us\n",
(int)my_col, ((float)total)/1000000, ((float)total)/iters);
fflush(stdout);
}
BARRIER();
MSG("done.");
gasnet_exit(0); /* for faster exit */
return 0;
}
int main(int argc, char **argv) {
const char *kind = "";
int pollers = 0;
int arg = 1;
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(htable, 1, TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
TEST_COLL_INIT();
#if GASNET_PAR
test_init("testbarrier", 1, "[-t] [-p polling_threads] (iters)\n"
" The -p option gives a number of polling threads to spawn (default is 0).\n"
" The -b option replaces barrier_notify calls with blocking barrier calls\n"
" The -t option replaces barrier_wait calls with looping on barrier_try");
#else
test_init("testbarrier", 1, "[-t] (iters)\n"
" The -b option replaces barrier_notify calls with blocking barrier calls\n"
" The -t option replaces barrier_wait calls with looping on barrier_try");
#endif
mynode = gasnet_mynode();
nodes = gasnet_nodes();
while (argc-arg >= 2) {
if (!strcmp(argv[arg], "-p")) {
#if GASNET_PAR
pollers = test_thread_limit(atoi(argv[arg+1])+1)-1;
arg += 2;
#else
if (gasnet_mynode() == 0) {
fprintf(stderr, "testbarrier %s\n", GASNET_CONFIG_STRING);
fprintf(stderr, "ERROR: The -p option is only available in the PAR configuration.\n");
fflush(NULL);
}
sleep(1);
gasnet_exit(1);
#endif
} else if (!strcmp(argv[arg], "-t")) {
do_try = 1;
kind = " polling";
arg += 1;
} else if (!strcmp(argv[arg], "-b")) {
do_block = 1;
kind = " blocking";
arg += 1;
}
}
if (argc-arg >= 1) iters = atoi(argv[arg]);
if (!iters) iters = 10000;
if (argc-arg >= 2) test_usage();
if (do_try && do_block) {
if (!mynode) fprintf(stderr, "ERROR: The -b and -t options are mutually exclusive.\n");
gasnet_exit(1);
}
#if !defined(GASNET_PAR)
MSG0("Running%s barrier test with %i iterations...\n",kind,iters);
#else
MSG0("Running%s barrier test with %i iterations and %i extra polling threads...\n",kind,iters,pollers);
TEST_SET_WAITMODE(pollers+1);
if (pollers)
test_createandjoin_pthreads(pollers+1,doTest,NULL,0);
else
#endif
doTest(NULL);
BARRIER();
MSG("done.");
gasnet_exit(0);
return 0;
}
int
main(int argc, char **argv)
{
int i;
const char *getopt_str;
int opt_p=0, opt_g=0, opt_a=0, opt_m=0;
#if TEST_MPI
init_test_mpi(&argc, &argv);
getopt_str = "pgamlvtdi:";
#else
getopt_str = "pgalvtdi:";
#endif
GASNET_Safe(gasnet_init(&argc, &argv));
GASNET_Safe(gasnet_attach(htable, HANDLER_TABLE_SIZE,
TEST_SEGSZ_REQUEST, TEST_MINHEAPOFFSET));
#if TEST_MPI
#define TEST_MPI_USAGE " -m use MPI calls \n"
#else
#define TEST_MPI_USAGE ""
#endif
#if GASNET_PAR
#define TEST_THREAD_USAGE " [<threads_per_node>]\n\n" \
"<threads_per_node> must be between 1 and "_STRINGIFY(TEST_MAXTHREADS)" \n"
#else
#define TEST_THREAD_USAGE "\n\n"
#endif
test_init("testthreads",0, "[ -pgalvtd ] [ -i <iters> ]"
TEST_THREAD_USAGE
"no options means run all tests with "_STRINGIFY(DEFAULT_ITERS)" iterations\n"
"options: \n"
" -p use puts \n"
" -g use gets \n"
" -a use Active Messages \n"
" -l use local Active Messages \n"
TEST_MPI_USAGE
" -v output information about actions taken \n"
" -t include AM handler actions with -v \n"
" -d dynamic thread creation stress test \n"
" -i <iters> use <iters> iterations per thread \n");
while ((i = getopt (argc, argv, getopt_str)) != EOF) {
switch (i) {
case 'p': opt_p = 1; break;
case 'g': opt_g = 1; break;
case 'a': opt_a = 1; break;
case 'm': opt_m = 1; break;
case 'l': AM_loopback = 1; break;
case 'i': iters = atoi(optarg); break;
case 'v': verbose = 1; break;
case 't': amtrace = 1; break;
case 'd': threadstress = 1; break;
default: test_usage();
}
}
if (opt_p) test_functions[functions_num++] = test_put;
if (opt_g) test_functions[functions_num++] = test_get;
if (opt_a) {
test_functions[functions_num++] = test_amshort;
test_functions[functions_num++] = test_ammedium;
test_functions[functions_num++] = test_amlong;
}
#if TEST_MPI
if (opt_m) test_functions[functions_num++] = test_mpi;
#endif
if (amtrace) verbose = 1;
/* Assume all test functions if no option is passed */
if (functions_num == 0) {
MSG("running all functions!");
memcpy(test_functions, test_functions_all,
sizeof(test_functions_all));
functions_num = NUM_FUNCTIONS;
}
argc -= optind;
if (argc > 1) test_usage();
else if (argc == 1) {
argv += optind;
threads_num = atoi(argv[0]);
}
if (threads_num > TEST_MAXTHREADS || threads_num < 1) {
printf("ERROR: Threads must be between 1 and %i\n",TEST_MAXTHREADS);
exit(EXIT_FAILURE);
}
/* limit sizes to a reasonable size */
#define LIMIT(sz) MIN(sz,4194304)
{ int sz = 0;
sizes[sz++] = LIMIT(gasnet_AMMaxMedium()-1);
sizes[sz++] = LIMIT(gasnet_AMMaxMedium());
sizes[sz++] = LIMIT(gasnet_AMMaxMedium()+1);
sizes[sz++] = LIMIT(gasnet_AMMaxLongRequest()-1);
sizes[sz++] = LIMIT(gasnet_AMMaxLongRequest());
sizes[sz++] = LIMIT(gasnet_AMMaxLongRequest()+1);
sizes[sz++] = LIMIT(gasnet_AMMaxLongReply()-1);
sizes[sz++] = LIMIT(gasnet_AMMaxLongReply());
sizes[sz++] = LIMIT(gasnet_AMMaxLongReply()+1);
assert(sizes[sz] == 0);
}
alloc_thread_data(threads_num);
#if TEST_MPI
attach_test_mpi();
#endif
#ifdef GASNET_PAR
if (threadstress) {
int spawniters = MAX(1,iters/threads_num);
int i;
MSG("Dynamic thread creation stress test, %d gasnet threads, (%d at a time)", spawniters*threads_num, threads_num);
iters = 10; /* enough iters to ensure we get thread registration */
for (i = 0; i < spawniters; i++) {
test_createandjoin_pthreads(threads_num, &threadmain, tt_thread_data, sizeof(threaddata_t));
TEST_PROGRESS_BAR(i, spawniters);
}
} else {
MSG("Forking %d gasnet threads and running %d iterations", threads_num, iters);
test_createandjoin_pthreads(threads_num, &threadmain, tt_thread_data, sizeof(threaddata_t));
}
#else /* for testmpi-seq and -parsync */
#ifdef GASNET_SEQ
MSG("Running with 1 thread/node for GASNET_SEQ mode");
#else
MSG("Running with 1 thread/node for GASNET_PARSYNC mode");
#endif
threadmain(tt_thread_data);
#endif
BARRIER();
free_thread_data();
MSG("Tests complete");
BARRIER();
gasnet_exit(0);
return 0;
}