void smp_coll_tune_barrier(smp_coll_t handle) {
int iter;
int i;
gasnett_tick_t start,stop;
int radix=2;
double best_time=1e20;
double time;
int barrier_iters=gasneti_getenv_int_withdefault("GASNET_COLL_TUNE_SMP_BARRIER_ITER", 1000, 0);
int root;
static int best_barrier_radix = 2;
static int best_barrier_routine = SMP_COLL_BARRIER_DISSEM_ATOMIC;
static int best_root = 0;
#if VERBOSE_TUNING
if(handle->MYTHREAD==0) fprintf(stderr, "starting autotuning of local barrier\n");
#endif
for(root=0; root<1; root++) {
if(handle->MYTHREAD==0 && VERBOSE_TUNING) fprintf(stderr, "ROOT: %d\n", root);
for(i=0; i<SMP_COLL_NUM_BARR_ROUTINES; i++) {
if(i==SMP_COLL_BARRIER_COND_VAR) continue;
if(handle->MYTHREAD==0 && VERBOSE_TUNING) fprintf(stderr, "\t routine: %d\n",(int)i);
for(radix=2; radix<=handle->THREADS; radix*=2) {
if(i==SMP_COLL_BARRIER_COND_VAR && radix>2) continue;
if(i==SMP_COLL_BARRIER_PTHREAD && radix>2) continue;
if(handle->MYTHREAD==0 && VERBOSE_TUNING) fprintf(stderr, "\t\t radix: %d\n",radix);
smp_coll_set_barrier_routine_with_root(handle, (smp_coll_barrier_routine_t)i, radix, root);
start = gasnett_ticks_now();
for(iter = 0; iter<barrier_iters; iter++) {
smp_coll_barrier(handle, 0);
}
stop = gasnett_ticks_now();
smp_coll_safe_barrier(handle, 0);
time = ((double) gasnett_ticks_to_ns(stop-start))/barrier_iters;
if(time < best_time && handle->MYTHREAD==0) {
best_barrier_radix=radix;
best_barrier_routine = i;
best_time = time;
best_root = root;
}
smp_coll_reset_all_flags(handle);
/*run a root agnostic barrier*/
smp_coll_safe_barrier(handle, 0);
}
}
}
#if VERBOSE_TUNING
if(handle->MYTHREAD==0) fprintf(stderr, "setting best barrier: routine: %d radix: %d root: %d time: %g ns\n", best_barrier_routine, best_barrier_radix, best_root, best_time);
#endif
smp_coll_set_barrier_routine_with_root(handle, (smp_coll_barrier_routine_t)best_barrier_routine, best_barrier_radix, best_root);
}
void run() {
unsigned i;
gasnett_tick_t start, end;
hashtable_create(TABLE_SIZE);
BARRIER();
start = gasnett_ticks_now();
for (i = 0; i < WRITES_PER_PROC; ++i) {
hashtable_insert(grt_id * WRITES_PER_PROC + i);
}
end = gasnett_ticks_now();
unsigned time = ((unsigned) gasnett_ticks_to_us(end - start));
printf("processor %u: insertion time=%f us\n",
grt_id, (double) time);
fflush(stdout);
grt_write(0, time, ×[grt_id]);
BARRIER();
/* sanity check */
for (i = 0; i < WRITES_PER_PROC; ++i) {
grt_word_t num = grt_id * WRITES_PER_PROC + i;
grt_bool_t found = hashtable_find(num);
if (!found) {
fprintf(stderr, "processor %d: expected to find %d\n",
grt_id, (int) num);
}
}
BARRIER();
if (grt_id == 0) {
time = 0, max_time = 0;
for (i = 0; i < grt_num_procs; ++i) {
gasnett_tick_t this_time = times[i];
time += this_time;
if (this_time >= max_time) max_time = this_time;
}
time_per_op = ((float) time) / NUM_WRITES;
printf("total CPU time=%f us\n", (double) time);
printf("time per operation=%f us\n", time_per_op);
printf("max time=%f us\n", (double) max_time);
}
BARRIER();
hashtable_destroy();
BARRIER();
}
void
test_sleep(threaddata_t *tdata)
{
unsigned usecs = (unsigned) sleep_min_us +
(rand() % (sleep_max_us - sleep_min_us));
ACTION_PRINTF("tid=%3d> sleeping %.3f millisecs", tdata->tid, usecs/1000.0);
{ uint64_t goal = gasnett_ticks_to_us(gasnett_ticks_now()) + usecs;
while (gasnett_ticks_to_us(gasnett_ticks_now()) < goal)
gasnett_sched_yield();
}
ACTION_PRINTF("tid=%3d> awaking", tdata->tid);
}
void run() {
unsigned i, time;
gasnett_tick_t start, end;
hashset_create(params[HASHSET_SIZE], params[ON_PTHREAD]);
BARRIER();
start = gasnett_ticks_now();
for (i = 0; i < MY_NUM_OPS; ++i) {
if (put_flags[i] == GRT_TRUE) {
hashset_insert(numbers[i]);
} else {
hashset_find(numbers[i]);
}
}
end = gasnett_ticks_now();
time = ((unsigned) gasnett_ticks_to_us(end - start));
//printf("processor %u: execution time=%f us\n",
// grt_id, (double) time);
fflush(stdout);
grt_write(0, time, ×[grt_id]);
BARRIER();
if (grt_id == 0) {
time = 0, max_time = 0;
for (i = 0; i < grt_num_procs; ++i) {
gasnett_tick_t this_time = times[i];
time += this_time;
if (this_time >= max_time) max_time = this_time;
}
time_per_op = ((float) time) / params[NUM_OPS];
printf("total CPU time=%f us\n", (double) time);
printf("time per operation=%f us\n", time_per_op);
printf("max time=%f us\n", (double) max_time);
}
BARRIER();
hashset_destroy();
BARRIER();
}
/* ------------------------------------------------------------------------------------ */
void doit1(void) {
GASNET_BEGIN_FUNCTION();
{
int i;
for (i=0; i<8; i++) handles[i] = GASNET_INVALID_HANDLE;
}
TEST_SECTION_BEGIN();
TIME_OPERATION("Tester overhead", {});
TIME_OPERATION("gasnett_ticks_now()",
{ timertemp = gasnett_ticks_now(); });
void run() {
unsigned i, time;
gasnett_tick_t start, end;
hash_map_create(params[HASHMAP_SIZE], (grt_bool_t) params[ON_PTHREAD]);
grt_barrier();
#ifdef LOCKS
grt_lock_state_t state;
#endif
for (i = 0; i < MY_NUM_OPS; ++i) {
grt_word_t key = keys[i], val = values[i];
#ifdef LOCKS
hash_t hash = compute_hash(key);
hash_map_lock(hash.proc, hash.offset, WRITE, &state);
#endif
hash_map_insert(key, val);
#ifdef LOCKS
hash_map_unlock(hash.proc, hash.offset);
#endif
}
BARRIER();
start = gasnett_ticks_now();
#ifdef LOCKS
grt_lock_state_t state1, state2;
#endif
for (i = 0; i < MY_NUM_OPS; ++i) {
unsigned idx = grt_random_next() * MY_NUM_OPS;
grt_word_t key1 = keys[i];
unsigned second_idx = grt_random_next() * MY_NUM_OPS;
grt_word_t key2 = keys[second_idx];
#ifdef LOCKS
lock(key1, key2, &state1, &state2);
#endif
grt_word_t val1, val2;
#ifndef LOCKS
#ifndef NOLOCKS
stm_start(grt_id);
#endif
#endif
grt_bool_t found1 = hash_map_find(key1, &val1);
grt_bool_t found2 = hash_map_find(key2, &val2);
hash_map_insert(key1, val2);
hash_map_insert(key2, val1);
#ifndef LOCKS
#ifndef NOLOCKS
stm_commit(grt_id);
#endif
#endif
#if LOCKS
unlock(key1, key2);
#endif
}
end = gasnett_ticks_now();
time = ((unsigned) gasnett_ticks_to_us(end - start));
printf("processor %u: execution time=%f us\n",
grt_id, (double) time);
fflush(stdout);
grt_write(0, time, ×[grt_id]);
BARRIER();
if (grt_id == 0) {
time = 0, max_time = 0;
for (i = 0; i < grt_num_procs; ++i) {
gasnett_tick_t this_time = times[i];
time += this_time;
if (this_time >= max_time) max_time = this_time;
}
time_per_op = ((float) time) / params[NUM_OPS];
printf("total CPU time=%f us\n", (double) time);
printf("time per operation=%f us\n", time_per_op);
printf("max time=%f us\n", (double) max_time);
}
BARRIER();
hash_map_destroy();
BARRIER();
}
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;
}
/* placed in a function to avoid excessive inlining */
gasnett_tick_t ticktime(void) { return gasnett_ticks_now(); }