int main(int argc, char *argv[])
{
unsigned long i;
gaspi_pointer_t _vptr;
gaspi_rank_t num_ranks, myrank;
gaspi_number_t qmax ;
gaspi_number_t queueSize;
gaspi_rank_t left_rank, right_rank;
const unsigned long N = (1 << 13);
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_proc_num(&num_ranks));
ASSERT (gaspi_proc_rank(&myrank));
ASSERT(gaspi_segment_create(0,
_2MB,
GASPI_GROUP_ALL,
GASPI_BLOCK,
GASPI_MEM_INITIALIZED));
ASSERT(gaspi_segment_ptr(0, &_vptr));
ASSERT (gaspi_queue_size_max(&qmax));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
left_rank = (myrank + num_ranks - 1 ) % num_ranks;
right_rank = (myrank + num_ranks + 1) % num_ranks;
ASSERT( gaspi_write(0, //seg
0, //local off
left_rank, //rank
0, //seg rem
0, //remote off
1, //size 32KB
0, //queue
GASPI_BLOCK));
ASSERT( gaspi_write(0, //seg
0, //local off
right_rank, //rank
0, //seg rem
0, //remote off
1, //size 32KB
0, //queue
GASPI_BLOCK));
ASSERT (gaspi_wait(0, GASPI_BLOCK));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, 5000));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
printf("Rank %d: Finish\n", myrank);
fflush(stdout);
return EXIT_SUCCESS;
}
void work(int tid)
{
gaspi_rank_t rankSend;
gaspi_offset_t localOff= 0;
gaspi_offset_t remOff = 0;
gaspi_number_t queueSize, qmax;
gaspi_size_t commSize ;
ASSERT (gaspi_queue_size_max(&qmax));
for(commSize= 1; commSize < _500MB; commSize*=2 )
for(rankSend = 0; rankSend < numranks; rankSend++)
{
gaspi_printf("thread %d rank to send: %d - %lu bytes\n", tid, rankSend, commSize);
gaspi_queue_size(1, &queueSize);
if (queueSize > qmax - 100)
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_write(0, localOff, rankSend, 0, remOff, commSize, 1, GASPI_BLOCK));
}
ASSERT (gaspi_wait(1, GASPI_BLOCK));
gaspi_threads_sync();
}
void work(int tid)
{
gaspi_rank_t rankSend;
gaspi_offset_t localOff = 81478066;
gaspi_offset_t remOff = 81478246;
gaspi_offset_t size = 1800;
gaspi_number_t queueSize, qmax;
ASSERT (gaspi_queue_size_max(&qmax));
for(rankSend = 0; rankSend < numranks; rankSend++)
{
gaspi_printf("thread %d rank to send: %d\n", tid, rankSend);
gaspi_queue_size(1, &queueSize);
if (queueSize > qmax - 24)
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_write(0, localOff, rankSend, 0, remOff, size, 1, GASPI_BLOCK));
}
ASSERT (gaspi_wait(1, GASPI_BLOCK));
gaspi_threads_sync();
}
int main(int argc, char *argv[])
{
gaspi_rank_t numranks, myrank;
gaspi_rank_t rankSend;
gaspi_size_t segSize;
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_proc_num(&numranks));
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_segment_create(0, _2GB, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED));
ASSERT( gaspi_segment_size(0, myrank, &segSize));
gaspi_printf("seg size %lu MB \n", segSize/1024/1024);
// ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
gaspi_offset_t localOff= 814780664;
gaspi_offset_t remOff = 81478246;
gaspi_offset_t size = 1800;
gaspi_number_t queueSize, qmax;
ASSERT (gaspi_queue_size_max(&qmax));
for(rankSend = 0; rankSend < numranks; rankSend++)
{
gaspi_printf("rank to send: %d\n", rankSend);
gaspi_queue_size(1, &queueSize);
if (queueSize > qmax - 24)
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_write(0, localOff, rankSend, 0, remOff, size, 1, GASPI_BLOCK));
}
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return EXIT_SUCCESS;
}
int main(int argc, char *argv[])
{
gaspi_rank_t numranks, myrank;
gaspi_rank_t rankSend;
gaspi_size_t segSize;
const gaspi_offset_t localOff= 0;
const gaspi_offset_t remOff = 0;
gaspi_number_t queueSize, qmax;
gaspi_size_t commSize ;
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_proc_num(&numranks));
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_segment_create(0, _8MB, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_queue_size_max(&qmax));
for(commSize= 1; commSize <= _8MB; commSize*=2 )
{
for(rankSend = 0; rankSend < numranks; rankSend++)
{
gaspi_printf("rank to send: %d - %lu bytes\n", rankSend, commSize);
gaspi_queue_size(1, &queueSize);
if (queueSize > qmax - 24)
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_write(0, localOff, rankSend, 0, remOff, commSize, 1, GASPI_BLOCK));
}
}
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return EXIT_SUCCESS;
}
int main (int argc, char *argv[])
{
gaspi_configuration_t config = { 0 };
//argc, argv, "", (1UL << 30)
gaspi_proc_init (config, GASPI_BLOCK); // 1 GiB DMA enabled memory per node
gaspi_rank_t iProc, nProc;
gaspi_proc_rank (&iProc);
gaspi_proc_num (&nProc);
void* temp_ptr;
gaspi_segment_ptr(GPI_SEGMENT, &temp_ptr);
int *mem = (int *) temp_ptr; // begin of DMA enabled memory
int *src = mem; // offset 0
int *dst = mem + nProc; // offset nProc * sizeof(int)
for (gaspi_rank_t p = 0; p < nProc; ++p)
{
src[p] = iProc * nProc + p;
const unsigned long locOff = p * sizeof (int);
const unsigned long remOff = (nProc + iProc) * sizeof (int);
gaspi_write(GPI_SEGMENT, locOff, p, GPI_SEGMENT, remOff, sizeof (int), 0, GASPI_BLOCK);
}
gaspi_wait (0, GASPI_BLOCK);
gaspi_barrier (GASPI_GROUP_ALL, GASPI_BLOCK);
dump (src, iProc, nProc, "src");
dump (dst, iProc, nProc, "dst");
gaspi_proc_term (GASPI_BLOCK);
return EXIT_SUCCESS;
}
void
send_global_msg_to_check_state(gaspi_state_vector_t health_vec, gaspi_rank_t *avoid_list)
{
int i, j;
int num_simultaneous_fail_checks = 1;
gaspi_timeout_t HEALTH_CHECK_TIMEOUT_TIME = GASPI_BLOCK;
gaspi_printf("Checking global health state\n");
/* in order to check multiple simultaneous fail, health check has to be performed multiple times */
for(j = 0 ; j < num_simultaneous_fail_checks; ++j )
{
for(i = 0; i < numprocs; ++i)
{
if(avoid_list[i] != 1)
{
ASSERT(gaspi_write(gm_seg_health_chk_array_id, myrank, i,
gm_seg_health_chk_array_id, myrank, sizeof(int),
queue_id, HEALTH_CHECK_TIMEOUT_TIME));
}
}
gaspi_wait(queue_id, HEALTH_CHECK_TIMEOUT_TIME);
ASSERT(gaspi_state_vec_get(health_vec));
/* adding the dead processes to avoid_list */
/* so that message for health test is not sent to them next time. */
for(i = 0; i < numprocs; ++i)
{
if(health_vec[i] == 1)
{
avoid_list[i] = 1;
}
}
}
print_health_vec(health_vec);
}
int main(int argc, char *argv[])
{
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
const unsigned long N = (1 << 13);
gaspi_rank_t P, myrank;
ASSERT (gaspi_proc_num(&P));
ASSERT (gaspi_proc_rank(&myrank));
gaspi_printf("P = %d N = %lu\n", P, N);
gaspi_printf("Seg size: %lu MB\n", MAX (_4GB, 2 * ((N/P) * N * 2 * sizeof (double)))/1024/1024);
if(gaspi_segment_create(0, _1GB,
GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED) != GASPI_SUCCESS){
gaspi_printf("Failed to create segment\n");
return -1;
}
gaspi_pointer_t _vptr;
if(gaspi_segment_ptr(0, &_vptr) != GASPI_SUCCESS)
printf("gaspi_segment_ptr failed\n");
gaspi_number_t qmax ;
ASSERT (gaspi_queue_size_max(&qmax));
gaspi_printf("Queue max: %lu\n", qmax);
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
int i;
gaspi_number_t queueSize;
int rankSend = (myrank + 1) % P;
gaspi_printf("rank to: %d\n", rankSend);
for (i = 0; i < 2 * N; i ++)
{
gaspi_queue_size(1, &queueSize);
if (queueSize > qmax - 24)
{
gaspi_return_t ret;
do
{
ret = gaspi_wait(1, GASPI_TEST);
assert (ret != GASPI_ERROR);
}
while(ret != GASPI_SUCCESS);
gaspi_queue_size(1, &queueSize);
assert(queueSize == 0);
}
ASSERT (gaspi_write(0, 4, rankSend, 0, 6, 32768, 1, GASPI_TEST));
}
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return EXIT_SUCCESS;
}
int main(int argc, char *argv[])
{
gaspi_rank_t nprocs, myrank, i;
int j, n;
gaspi_rank_t *avoid_list;
gaspi_group_t survivors;
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT(gaspi_proc_num(&nprocs));
ASSERT(gaspi_proc_rank(&myrank));
ASSERT(gaspi_segment_create(0,
_4MB,
GASPI_GROUP_ALL,
GASPI_BLOCK,
GASPI_MEM_INITIALIZED));
avoid_list = (gaspi_rank_t *) malloc(nprocs * sizeof(gaspi_rank_t));
assert (avoid_list != NULL);
memset(avoid_list, 0, nprocs * sizeof(gaspi_rank_t));
gaspi_state_vector_t vec = (gaspi_state_vector_t) malloc(nprocs);
ASSERT(gaspi_state_vec_get(vec));
//check that everyone is healthy
for(i = 0; i < nprocs; i++)
{
assert(vec[i] == 0);
}
//sync
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
//now last rank dies
if(myrank == nprocs - 1)
exit(-1);
else
{
//create group of survivors
ASSERT(gaspi_group_create(&survivors));
for(i = 0; i < nprocs - 1; i++)
ASSERT(gaspi_group_add(survivors, i));
ASSERT(gaspi_group_commit(survivors, GASPI_BLOCK));
gaspi_printf("Done with groups\n");
sleep(2);
}
//the others communicate
gaspi_return_t retval;
for(j = 0; j < 10; j++)
{
gaspi_printf("Iteration %d\n", j);
for(i = 0; i < nprocs; i++)
{
if( avoid_list[i] != 1 )
ASSERT(gaspi_write(0, 0, i,
0, 0, sizeof(int),
0,
GASPI_BLOCK));
}
retval = gaspi_wait(0, GASPI_BLOCK);
//problem found -> recover
if(retval != GASPI_SUCCESS)
{
ASSERT(gaspi_state_vec_get(vec));
for(n = 0; n < nprocs; n++)
{
if(vec[n] != GASPI_STATE_HEALTHY)
{
gaspi_printf("Problem with node %d detected\n", n);
assert(n == (nprocs - 1));
ASSERT(recover());
avoid_list[n] = 1;
}
}
}
}
ASSERT (gaspi_barrier(survivors, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
gaspi_printf("exiting\n");
return EXIT_SUCCESS;
}
int main(int argc, char *argv[])
{
int i, iter;
int ret = 0;
gaspi_rank_t myrank, numranks;
gaspi_size_t mem_size = 0UL, j;
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_proc_num(&numranks));
if( numranks < 2 )
{
return EXIT_SUCCESS;
}
mem_size = 2 * SLOT_SIZE * (numranks - 1);
if(myrank == 0)
{
printf("Mem size: %lu (%.2f MB)\nProcs: %u Max Slot size %lu Iterations %d\n",
mem_size,
mem_size * 1.0f / 1024/ 1024,
numranks,
(gaspi_size_t) SLOT_SIZE,
MAX_ITERATIONS);
#ifdef WITH_SYNC
printf("Using notifications only\n");
#endif
}
ASSERT (gaspi_segment_create(0, mem_size, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED));
gaspi_pointer_t _vptr;
ASSERT (gaspi_segment_ptr(0, &_vptr));
float *mptr = (float *) _vptr;
//generate random
srand((unsigned)time(0));
srand48((unsigned) time(0));
gaspi_size_t cur_slot_size = SLOT_SIZE;
for(cur_slot_size = SLOT_SIZE; cur_slot_size >= sizeof(float); cur_slot_size/=2)
{
if(myrank == 0)
printf("===== Slot Size %lu ====\n", cur_slot_size);
for(iter = 0; iter < MAX_ITERATIONS; iter++)
{
if(myrank == 0)
printf("iteration %3d... ", iter);
ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
/* fill slots with randoms */
for(j = 0; j < (mem_size / sizeof(float) / 2); j++)
{
mptr[j]= drand48() + (myrank*1.0);
}
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
gaspi_offset_t offset_in = 0, offset_out = mem_size / 2;
/* rank 0 write to all others */
if(myrank == 0)
{
for (i = 1; i < numranks; i++)
{
offset_in = (i - 1) * cur_slot_size;
#ifdef WITH_SYNC
ASSERT (gaspi_write_notify(0, offset_in, i,
0, 0, cur_slot_size,
0, 1,
0, GASPI_BLOCK));
#else
ASSERT (gaspi_write(0, offset_in, i,
0, 0, cur_slot_size,
0, GASPI_BLOCK));
#endif
}
ASSERT(gaspi_wait(0, GASPI_BLOCK));
}
#ifdef WITH_SYNC
else
{
gaspi_notification_id_t id;
gaspi_notification_t val;
ASSERT(gaspi_notify_waitsome(0, 0, 1, &id, GASPI_BLOCK));
ASSERT(gaspi_notify_reset(0, id, &val));
assert(val == 1);
}
#endif
#ifndef WITH_SYNC
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
#endif
/* other ranks all write back to 0 */
if(myrank != 0)
{
offset_in = 0;
offset_out = (mem_size / 2) + (cur_slot_size * (myrank - 1));
#ifdef WITH_SYNC
ASSERT (gaspi_write_notify(0, offset_in, 0,
0, offset_out, cur_slot_size,
myrank, 1,
0, GASPI_BLOCK));
#else
ASSERT (gaspi_write(0, offset_in, i,
0, offset_out, cur_slot_size,
0, GASPI_BLOCK));
#endif
ASSERT(gaspi_wait(0, GASPI_BLOCK));
}
#ifdef WITH_SYNC
else
{
gaspi_notification_id_t id;
gaspi_notification_t val;
int notification_counter = 0;
do
{
ASSERT(gaspi_notify_waitsome(0, 1, numranks - 1, &id, GASPI_BLOCK));
ASSERT(gaspi_notify_reset(0, id, &val));
assert(val == 1);
notification_counter++;
}
while( notification_counter < numranks - 1);
}
#endif
#ifndef WITH_SYNC
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
#endif
if(myrank == 0)
{
/* check correctness */
float *in = (float *) _vptr;
float *out = (float *) ((char *) _vptr + mem_size / 2);
const gaspi_size_t total_elems = (cur_slot_size * (numranks - 1) / sizeof(float));
for(j = 0; j < total_elems; j++)
{
if(in[j] != out[j])
{
printf("Different values at pos %lu: %f %f (iterations %d)\n", j, in[j], out[j], iter);
ret = -1;
goto end;
}
}
printf("All fine!\n");
}
}
}
end:
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return ret;
}
int main(int argc, char *argv[])
{
int k = 0;
int ret = 0;
unsigned long j;
const gaspi_size_t size = 4096;
const gaspi_size_t memSize = _4GB;
gaspi_offset_t offset_write = 0;
gaspi_offset_t offset_read = _2GB;
gaspi_offset_t offset_check = 3221225472;
gaspi_number_t qmax ;
gaspi_number_t queueSize;
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_queue_size_max(&qmax));
ASSERT (gaspi_segment_create(0, memSize, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED));
gaspi_pointer_t _vptr;
ASSERT (gaspi_segment_ptr(0, &_vptr));
/* get memory area pointer */
float *mptr_f = (float *) _vptr;
char *mptr_c = (char *) _vptr;
gaspi_rank_t myrank, highestnode;
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_proc_num(&highestnode));
while(k <= RUNS)
{
//generate random
srand((unsigned)time(0));
srand48((unsigned) time(0));
ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
//clean
for(j = 0; j < memSize; j++)
mptr_c[j]= 0;
/* fill randoms up to 1GB */
for(j = 0; j < (GB / sizeof(float)); j++)
{
mptr_f[j]= drand48() + (myrank * 1.0);
}
#ifdef DEBUG
gaspi_printf("random value in pos 0 %f\n", mptr_f[0]);
#endif
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
gaspi_printf("\n....Running iteration %d of %d...\n",k, RUNS);
const unsigned long packets = (GB / size);
for(j = 0; j < packets; j++)
{
ASSERT(gaspi_queue_size(0, &queueSize));
if (queueSize > qmax - 24)
{
ASSERT(gaspi_wait(0, GASPI_BLOCK));
}
ASSERT (gaspi_write(0, offset_write, (myrank + 1) % highestnode,
0, offset_read, size,
0, GASPI_BLOCK));
offset_write += size;
offset_read += size;
}
offset_write=0;
offset_read = _2GB;
#ifdef DEBUG
gaspi_printf("%d bytes written!\n", packets * size);
#endif
/* notify remote that data is written */
ASSERT (gaspi_notify( 0, (myrank + 1) % highestnode, 0, 1, 0, GASPI_BLOCK));
gaspi_notification_id_t recv_id;
ASSERT(gaspi_notify_waitsome(0, 0, 1, &recv_id, GASPI_BLOCK));
assert(recv_id == 0);
gaspi_notification_t notification_val;
ASSERT( gaspi_notify_reset(0, recv_id, ¬ification_val));
/* notify remote that data has arrived */
ASSERT (gaspi_notify( 0, (myrank + highestnode - 1) % highestnode, 1, 1, 0, GASPI_BLOCK));
gaspi_notification_id_t ack_id;
ASSERT(gaspi_notify_waitsome(0, 1, 1, &ack_id, GASPI_BLOCK));
assert(ack_id == 1);
ASSERT( gaspi_notify_reset(0, ack_id, ¬ification_val));
/* check if data was written successfully */
ASSERT (gaspi_read(0, offset_check, (myrank + 1) % highestnode,
0, offset_read, GB / 2,
0, GASPI_BLOCK));
ASSERT (gaspi_read(0, offset_check + (GB / 2), (myrank + 1) % highestnode,
0, offset_read + (GB / 2), GB / 2,
0, GASPI_BLOCK));
ASSERT (gaspi_wait(0, GASPI_BLOCK));
#ifdef DEBUG
gaspi_printf("Values %f %f %f \n", mptr_f[0], mptr_f[offset_read / sizeof(float)], mptr_f[offset_check / sizeof(float)]);
#endif
j = 0;
while(j < GB / sizeof(float) )
{
if(mptr_f[j] != mptr_f[offset_check / sizeof(float) + j]){
gaspi_printf("value incorrect %f-%f at %d \n",
mptr_f[j],
mptr_f[offset_check / sizeof(float) + j],
j);
ret = -1;
goto out;
}
j++;
}
#ifdef DEBUG
gaspi_printf("Check!\n");
#endif
k++;
}
out:
gaspi_printf("Waiting to finish...\n");
ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return ret;
}
int main(int argc, char *argv[])
{
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
gaspi_rank_t numranks, myrank;
ASSERT (gaspi_proc_num(&numranks));
ASSERT (gaspi_proc_rank(&myrank));
int rankSend = (myrank + 1) % numranks;
ASSERT(gaspi_segment_create(0, _1MB, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED));
gaspi_size_t segSize;
ASSERT( gaspi_segment_size(0, myrank, &segSize));
unsigned char * pGlbMem;
gaspi_pointer_t _vptr;
ASSERT(gaspi_segment_ptr(0, &_vptr));
pGlbMem = ( unsigned char *) _vptr;
gaspi_number_t queueSize, qmax ;
ASSERT (gaspi_queue_size_max(&qmax));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
const unsigned long localOff = 0;
const unsigned long remOff = 0;
/* write_notify */
do
{
ASSERT(gaspi_write_notify(0, localOff, rankSend,
0, remOff, 1,
(gaspi_notification_id_t) myrank, 1,
1, GASPI_BLOCK));
gaspi_queue_size(1, &queueSize);
}
while(queueSize < qmax);
EXPECT_FAIL_WITH(gaspi_write_notify(0, localOff, rankSend,
0, remOff, 1,
(gaspi_notification_id_t) myrank, 1,
1, GASPI_BLOCK),
GASPI_QUEUE_FULL);
ASSERT (gaspi_wait(1, GASPI_BLOCK));
/* write */
do
{
ASSERT(gaspi_write(0, localOff, rankSend,
0, remOff, 1,
1, GASPI_BLOCK));
gaspi_queue_size(1, &queueSize);
}
while(queueSize < qmax);
EXPECT_FAIL_WITH(gaspi_write(0, localOff, rankSend,
0, remOff, 1,
1, GASPI_BLOCK),
GASPI_QUEUE_FULL);
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT(gaspi_write(0, localOff, rankSend,
0, remOff, 1,
1, GASPI_BLOCK));
/* write + write_notify */
do
{
ASSERT(gaspi_write(0, localOff, rankSend,
0, remOff, 1,
1, GASPI_BLOCK));
gaspi_queue_size(1, &queueSize);
}
while(queueSize < qmax - 1);
EXPECT_FAIL_WITH(gaspi_write_notify(0, localOff, rankSend,
0, remOff, 1,
(gaspi_notification_id_t) myrank, 1,
1, GASPI_BLOCK),
GASPI_QUEUE_FULL);
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT(gaspi_write_notify(0, localOff, rankSend,
0, remOff, 1,
(gaspi_notification_id_t) myrank, 1,
1, GASPI_BLOCK));
/* read */
do
{
ASSERT(gaspi_read(0, localOff, rankSend,
0, remOff, 1,
1, GASPI_BLOCK));
gaspi_queue_size(1, &queueSize);
}
while(queueSize < qmax);
EXPECT_FAIL_WITH(gaspi_read(0, localOff, rankSend,
0, remOff, 1,
1, GASPI_BLOCK),
GASPI_QUEUE_FULL);
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT(gaspi_read(0, localOff, rankSend,
0, remOff, 1,
1, GASPI_BLOCK));
/* write_list_notify */
{
const gaspi_number_t nListElems = 255;
gaspi_number_t n;
gaspi_segment_id_t localSegs[nListElems];
gaspi_offset_t localOffs[nListElems];
const gaspi_rank_t rank2send = (myrank + 1) % numranks;
gaspi_segment_id_t remSegs[nListElems];
gaspi_offset_t remOffs[nListElems];
gaspi_size_t sizes[nListElems];
const unsigned int bytes = sizeof(int);
gaspi_offset_t initLocOff = 0;
gaspi_offset_t initRemOff = (bytes * nListElems + 64);
for(n = 0; n < nListElems; n++)
{
sizes[n] = bytes;
localSegs[n] = 0;
localOffs[n] = initLocOff;
initLocOff += bytes;
remSegs[n] = 0;
remOffs[n] = initRemOff;
initRemOff += bytes;
}
do
{
ASSERT( gaspi_write_list_notify( nListElems,
localSegs, localOffs, rank2send,
remSegs, remOffs, sizes,
0, myrank, 1,
0, GASPI_BLOCK));
gaspi_queue_size(0, &queueSize);
}
while(queueSize < qmax);
EXPECT_FAIL_WITH( gaspi_write_list_notify( nListElems,
localSegs, localOffs, rank2send,
remSegs, remOffs, sizes,
0, myrank, 1,
0, GASPI_BLOCK),
GASPI_QUEUE_FULL);
ASSERT (gaspi_wait(0, GASPI_BLOCK));
ASSERT( gaspi_write_list_notify( nListElems,
localSegs, localOffs, rank2send,
remSegs, remOffs, sizes,
0, myrank, 1,
0, GASPI_BLOCK));
}
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return EXIT_SUCCESS;
}
int main(int argc, char* argv[])
{
int i, j;
gaspi_number_t gsize;
int comm_state = WORKING;
int num_failures = 0;
int timesteps = 0;
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_proc_num(&numprocs));
read_params(argc, argv, ×teps, &numprocs_idle);
numprocs_working = numprocs - numprocs_idle;
numprocs_working_and_idle = numprocs_working + numprocs_idle;
gaspi_rank_t *comm_main_ranks = malloc( numprocs_idle * sizeof(gaspi_rank_t));
init_array_2(comm_main_ranks, numprocs_working);
/* contains info of all processes:
which are working(0), broken(1) and idle(2).
keeps updated all the time(iterations) */
int * status_processes = (int *) malloc(numprocs * sizeof(int));
init_array_3(status_processes, numprocs, WORKING);
for(i = numprocs-1, j=0; j < numprocs_idle;--i,++j)
{
status_processes[i] = IDLE; // putting last processes to IDLE
}
// ===== GASPI group creation =====
if(status_processes[myrank]==WORKING)
{
ASSERT(gaspi_group_create(&COMM_MAIN));
gaspi_number_t i;
for(i=0; i<numprocs; i++)
{
if(status_processes[i]==WORKING)
{
ASSERT(gaspi_group_add(COMM_MAIN, i));
ASSERT(gaspi_group_size(COMM_MAIN, &gsize));
}
}
ASSERT(gaspi_group_ranks (COMM_MAIN, comm_main_ranks));
ASSERT(gaspi_group_commit (COMM_MAIN, GASPI_BLOCK));
}
/* ====== Init a SYNC FLAGS Segment ====== */
/* used to communicate the WORKING, BROKEN, or FINISHED_WORK status between the working and idle processes. */
gaspi_size_t SYNC_global_mem_size;
SYNC_global_mem_size = numprocs * sizeof(int);
gaspi_pointer_t gm_ptr_sync=NULL;
ASSERT(init_segment (gm_seg_sync_flags_id, SYNC_global_mem_size));
ASSERT(gaspi_segment_ptr (gm_seg_sync_flags_id, &gm_ptr_sync));
int * sync_flags = (int *) gm_ptr_sync;
init_array_3(sync_flags, numprocs, WORKING);
/* ====== Init a health check write FLAGS Segment ====== */
/* This array is used to send the gaspi_write message write before health_chk routine,
which will then update the gaspi internal health vector */
gaspi_size_t health_chk_global_mem_size;
health_chk_global_mem_size = numprocs*sizeof(int);
gaspi_pointer_t gm_ptr_health_chk=NULL;
ASSERT(init_segment (gm_seg_health_chk_array_id, health_chk_global_mem_size));
ASSERT(gaspi_segment_ptr (gm_seg_health_chk_array_id, &gm_ptr_health_chk));
gaspi_state_vector_t health_vec = (gaspi_state_vector_t) malloc(numprocs);
ASSERT(gaspi_state_vec_get(health_vec));
gaspi_rank_t * avoid_list= (gaspi_rank_t *) malloc(numprocs * sizeof(gaspi_rank_t));
for(i = 0;i < numprocs; ++i)
avoid_list[i] = (gaspi_rank_t) 0;
gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK);
/* ===== TIME-STEP LOOP ===== */
if(status_processes[myrank]==IDLE)
{
/* IDLE processes remain in this loop */
while(1)
{
gaspi_printf("%d.", myrank);
if(sync_flags[0] == WORKING)
{
/* NO FAILURE REPORTED */
usleep(1000000);
}
if(sync_flags[0] == BROKEN)
{
/* FAILURE REPORTED */
gaspi_printf("myrank: %d Broken reported\n", myrank);
comm_state=BROKEN;
break;
}
if(sync_flags[0] == WORKFINISHED)
{
/* WORKFINISHED REPORTED */
gaspi_printf("myrank: %d WorkFinished reported\n", myrank);
comm_state = WORKFINISHED;
break;
}
}
}
int time_step;
for(time_step=1; time_step <= timesteps && comm_state!=WORKFINISHED; time_step++)
{
gaspi_printf("== time_step: %d ==\n", time_step);
if(comm_state==WORKING && status_processes[myrank]==WORKING)
{
gaspi_barrier(COMM_MAIN, GASPI_TIMEOUT_TIME);
sleep(1); // NOTE: this is the work section.
if(time_step == 5 && myrank== 1)
{
exit (-1);
}
}
if(time_step<timesteps )
{
send_global_msg_to_check_state(health_vec, avoid_list);
num_failures = check_comm_health(status_processes, health_vec);
gaspi_printf("%d NUM_FAILURES at timestep %d = %d\n", myrank, time_step, num_failures);
if( num_failures != 0 )
{
rescue_process = numprocs_working;
if(myrank==0)
{
// message the IDLE process
sync_flags[0]=BROKEN;
for(i = 0 ; i < num_failures ; ++i)
{
/* TODO: multiple failures at the same time. */
gaspi_printf("messaging rescue_process: %d\n", rescue_process);
ASSERT(gaspi_write(gm_seg_sync_flags_id, 0, rescue_process, gm_seg_sync_flags_id, 0, sizeof(int), 0, GASPI_BLOCK));
rescue_process++;
}
}
if(myrank==0 || myrank==rescue_process)
gaspi_printf("%d REPAIRING COMM_MAIN FLAG 1\n", myrank);
update_status_processes_array(status_processes, health_vec);
numprocs_working_and_idle = refresh_numprocs_working_and_idle(status_processes);
if(myrank != rescue_process)
{
ASSERT(gaspi_group_delete(COMM_MAIN));
ASSERT(recover());
}
ASSERT(gaspi_group_create(&COMM_MAIN_NEW));
for(i = 0; i < numprocs; i++)
{
if(status_processes[i]==WORKING)
{
ASSERT(gaspi_group_add(COMM_MAIN_NEW, i));
ASSERT(gaspi_group_size(COMM_MAIN_NEW, &gsize));
if(gsize == numprocs_working)
break;
}
}
gaspi_printf("%d: COMM_MAIN_NEW size is: %hi\n", myrank, gsize);
ASSERT(gaspi_group_commit (COMM_MAIN_NEW, GASPI_BLOCK));
init_array_2(comm_main_ranks, numprocs_working);
ASSERT(gaspi_group_ranks (COMM_MAIN_NEW, comm_main_ranks));
gaspi_printf("printing group_ranks_main: \n");
gaspi_printf_array(comm_main_ranks, numprocs_working);
comm_state = WORKING;
gaspi_printf("%d REPAIRING COMM_MAIN_NEW FLAG 2\n", myrank);
if(status_processes[myrank] == WORKING)
{
ASSERT(gaspi_barrier(COMM_MAIN_NEW, GASPI_BLOCK));
ASSERT(gaspi_barrier(COMM_MAIN_NEW, GASPI_BLOCK));
}
/* set things to work again */
COMM_MAIN = COMM_MAIN_NEW;
time_step = 5;
}
}
}
if(myrank == 0)
{
gaspi_printf("finished successfully\n");
}
gaspi_proc_term(10000);
return EXIT_SUCCESS;
}
int main(int argc, char *argv[])
{
int j,i,k=0;
int ret=0;
const gaspi_size_t size=4096;//4k
const gaspi_size_t memSize = 4294967296; //4GB
gaspi_offset_t offset_write=0, offset_read = memSize / 2, offset_check = 3221225472 ;
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_segment_create(0, memSize, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED));
gaspi_pointer_t _vptr;
ASSERT (gaspi_segment_ptr(0, &_vptr));
/* get memory area pointer */
#ifdef FLOAT
float *mptr = (float *) _vptr;
#else
int *mptr = (int *) _vptr;
#endif
gaspi_rank_t myrank, highestnode;
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_proc_num(&highestnode));
while(k <= RUNS)
{
//generate random
srand((unsigned)time(0));
#ifdef FLOAT
srand48((unsigned) time(0));
#endif
//clean
for(j = 0; j < (memSize / 4); j++)
mptr[j]= 0;
//fill randoms up to 1GB
for(j = 0; j < (memSize / 16); j++)
{
#ifdef FLOAT
mptr[j]= drand48() + (myrank*1.0);
#else
mptr[j]= rand() + myrank;
#endif
}
#ifdef DEBUG
#ifdef FLOAT
gaspi_printf("random value in pos 0 %f\n", mptr[0]);
#else
gaspi_printf("random value in pos 0 %d\n", mptr[0]);
#endif
#endif //DEBUG
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
gaspi_printf("\n....Running iteration %d of %d...\n",k, RUNS);
for(i = 0; i < ITERATIONS; i++)
{
for(j = 0; j < ITERATIONS; j++)
{
ASSERT (gaspi_write(0, offset_write, (myrank + 1) % highestnode,
0, offset_read, size, 0, GASPI_BLOCK));
offset_write += size;
offset_read += size;
}
ASSERT (gaspi_wait(0, GASPI_BLOCK));
}
#ifdef DEBUG
gaspi_printf("%d bytes written!\n", ITERATIONS * ITERATIONS * size);
#endif
//check if data was written successfully
ASSERT (gaspi_read(0, offset_check, (myrank + 1) % highestnode,
0, memSize/2, GB, 0, GASPI_BLOCK));
ASSERT (gaspi_wait(0, GASPI_BLOCK));
#ifdef DEBUG
gaspi_printf("%d bytes read!\n",GB);
#endif
j=0;
#ifdef DEBUG
#ifdef FLOAT
gaspi_printf("Values %f %f %f \n", mptr[0], mptr[memSize/8], mptr[offset_check/4]);
#else
gaspi_printf("Values %d %d %d \n", mptr[0], mptr[memSize/8], mptr[offset_check/4]);
#endif
#endif//DEBUG
while(j < GB / 4 )
{
if(mptr[j] != mptr[offset_check / 4 + j]){
#ifdef FLOAT
gaspi_printf("value incorrect %f-%f at %d \n",mptr[j],mptr[offset_check / 4],j);
#else
gaspi_printf("value incorrect %d-%d at %d \n",mptr[j],mptr[offset_check / 4],j);
#endif
ret = -1;
goto out;
}
j++;
}
offset_write=0;
offset_read = memSize / 2;
#ifdef DEBUG
gaspi_printf("Check!\n");
#endif
k++;
}
out:
gaspi_printf("Waiting to finish...\n");
ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return ret;
}
int main(int argc, char *argv[])
{
gaspi_rank_t numranks, myrank;
gaspi_rank_t rankSend;
gaspi_size_t segSize;
const gaspi_offset_t localOff_r= 0;
const gaspi_offset_t remOff_r = 0;
const gaspi_offset_t localOff_w = _128MB / 2 ;
const gaspi_offset_t remOff_w = _128MB / 2;
gaspi_number_t queueSize, qmax;
const gaspi_size_t commSize = _8MB;
int i;
gaspi_gpu_t gpus[8];
gaspi_gpu_num nGPUs;
TSUITE_INIT(argc, argv);
ASSERT (gaspi_proc_init(GASPI_BLOCK));
ASSERT (gaspi_proc_num(&numranks));
ASSERT (gaspi_proc_rank(&myrank));
ASSERT (gaspi_init_GPUs());
ASSERT (gaspi_number_of_GPUs(&nGPUs));
ASSERT (gaspi_GPU_ids(gpus));
ASSERT (gaspi_segment_create(0, _128MB, GASPI_GROUP_ALL, GASPI_BLOCK, GASPI_MEM_INITIALIZED|GASPI_MEM_GPU));
ASSERT( gaspi_segment_size(0, myrank, &segSize));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_queue_size_max(&qmax));
for(i = 0; i < 100; i++ )
{
for(rankSend = 0; rankSend < numranks; rankSend++)
{
if(rankSend == myrank)
continue;
gaspi_printf("partner rank: %d - %lu bytes (%d)\n", rankSend, commSize, i);
ASSERT (gaspi_queue_size(1, &queueSize));
if (queueSize > qmax - 24)
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_read(0, localOff_r, rankSend, 0, remOff_r, commSize, 1, GASPI_BLOCK));
}
}
for(i = 0; i < 100; i++ )
{
for(rankSend = 0; rankSend < numranks; rankSend++)
{
if(rankSend == myrank)
continue;
ASSERT (gaspi_queue_size(1, &queueSize));
if (queueSize > qmax - 24)
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_write(0, localOff_r, rankSend, 0, remOff_r, commSize, 1, GASPI_BLOCK));
}
}
ASSERT (gaspi_wait(1, GASPI_BLOCK));
ASSERT (gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
ASSERT (gaspi_proc_term(GASPI_BLOCK));
return EXIT_SUCCESS;
}
int
main (int argc, char *argv[])
{
int i, j, t;
gaspi_rank_t myrank;
char *ptr0;
//on numa architectures you have to map this process to the numa node where nic is installed
if (start_bench (2) != 0)
{
printf ("Initialization failed\n");
exit (-1);
}
// BENCH //
gaspi_proc_rank (&myrank);
if (gaspi_segment_ptr (0, (void **) &ptr0) != GASPI_SUCCESS)
{
printf ("gaspi_segment_ptr failed !\n");
exit (-1);
}
gaspi_float cpu_freq;
gaspi_cpu_frequency(&cpu_freq);
if (myrank < 2)
{
if(myrank == 0)
{
printf("-----------------------------------\n");
printf ("%12s\t%5s\n", "Bytes", "Lat(usecs)");
printf("-----------------------------------\n");
}
int bytes = 2;
volatile char *postBuf = (volatile char *) ptr0;
for (i = 1; i < 24; i++)
{
volatile char *pollBuf = (volatile char *) (ptr0 + ( 2 * bytes -1 ));
int rcnt = 0;
int cnt = 0;
gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK);
for (j = 0; j < ITERATIONS; j++)
{
if (rcnt < ITERATIONS && !(cnt < 1 && myrank == 1))
{
rcnt++;
while (*pollBuf != (char) rcnt)
{
#ifdef MIC
_mm_delay_32(32);
#else
_mm_pause();
#endif
}
}
stamp[j] = get_mcycles ();
postBuf[bytes - 1] = (char) ++cnt;
gaspi_write (0, 0, myrank ^ 0x1,
0, bytes, bytes,
0, GASPI_BLOCK);
gaspi_wait (0, GASPI_BLOCK);
}
for (t = 0; t < (ITERATIONS - 1); t++)
delta[t] = stamp[t + 1] - stamp[t];
qsort (delta, (ITERATIONS - 1), sizeof *delta, mcycles_compare);
const double div = 1.0 / cpu_freq;
const double ts = (double) delta[ITERATIONS / 2] * div * 0.5;
if(myrank == 0)
printf ("%12d\t%4.2f\n", bytes, ts);
bytes <<= 1;
}
}
end_bench ();
return 0;
}
int main(int argc, char *argv[])
{
int i;
gaspi_rank_t rank, nprocs;
gaspi_notification_id_t id;
const int num_elems = 1024;
TSUITE_INIT( argc, argv );
ASSERT( gaspi_proc_init(GASPI_BLOCK) );
ASSERT( gaspi_proc_num(&nprocs) );
ASSERT( gaspi_proc_rank(&rank) );
const gaspi_rank_t left = (rank + nprocs - 1 ) % nprocs;
const gaspi_rank_t right = (rank + nprocs + 1) % nprocs;
/* Create and fill buffer */
int * const buf = (int *) malloc(num_elems * sizeof(int));
assert( buf != NULL);
for (i = 0; i < num_elems; i++)
{
buf[i] = rank;
}
ASSERT( gaspi_segment_use( 0, buf, num_elems * sizeof(int),
GASPI_GROUP_ALL, GASPI_BLOCK,
0) );
ASSERT( gaspi_segment_create( 1, num_elems * sizeof(int),
GASPI_GROUP_ALL, GASPI_BLOCK,
GASPI_MEM_INITIALIZED) );
ASSERT(gaspi_barrier(GASPI_GROUP_ALL, GASPI_BLOCK));
/* write data to neighbour */
ASSERT( gaspi_write( 0, 0, right,
1, 0, num_elems * sizeof(int),
0, GASPI_BLOCK) );
ASSERT( gaspi_notify( 1, right, 0, 1, 0, GASPI_BLOCK ) );
ASSERT( gaspi_notify_waitsome( 1, 0, 1, &id, GASPI_BLOCK ) );
ASSERT( gaspi_wait( 0, GASPI_BLOCK ) );
/* Check data */
gaspi_pointer_t seg1_ptr;
ASSERT( gaspi_segment_ptr( 1, &seg1_ptr ) );
int * recv_buf = (int *) seg1_ptr;
for (i = 0; i < num_elems; i++)
{
assert(recv_buf[i] == left);
}
ASSERT( gaspi_segment_delete(0));
ASSERT( gaspi_segment_delete(1));
for (i = 0; i < num_elems; i++)
{
assert(buf[i] == rank);
}
ASSERT( gaspi_barrier( GASPI_GROUP_ALL, GASPI_BLOCK ) );
ASSERT( gaspi_proc_term( GASPI_BLOCK ) );
return EXIT_SUCCESS;
}
gaspi_return_t
pgaspi_gpu_write(const gaspi_segment_id_t segment_id_local,
const gaspi_offset_t offset_local,
const gaspi_rank_t rank,
const gaspi_segment_id_t segment_id_remote,
const gaspi_offset_t offset_remote,
const gaspi_size_t size,
const gaspi_queue_id_t queue,
const gaspi_timeout_t timeout_ms)
{
if( glb_gaspi_ctx.rrmd[segment_id_local][glb_gaspi_ctx.rank].cudaDevId < 0 ||
size <= GASPI_GPU_DIRECT_MAX )
{
return gaspi_write(segment_id_local, offset_local, rank, segment_id_remote, offset_remote, size, queue, timeout_ms);
}
if(lock_gaspi_tout (&glb_gaspi_ctx.lockC[queue], timeout_ms))
return GASPI_TIMEOUT;
char* host_ptr = (char*)(glb_gaspi_ctx.rrmd[segment_id_local][glb_gaspi_ctx.rank].host_ptr + NOTIFY_OFFSET + offset_local);
char* device_ptr = (char*)(glb_gaspi_ctx.rrmd[segment_id_local][glb_gaspi_ctx.rank].addr + offset_local);
gaspi_gpu* agpu = _gaspi_find_gpu(glb_gaspi_ctx.rrmd[segment_id_local][glb_gaspi_ctx.rank].cudaDevId);
if( !agpu )
{
gaspi_print_error("No GPU found or not initialized (gaspi_init_GPUs).");
return GASPI_ERROR;
}
int size_left = size;
int copy_size = 0;
int gpu_offset = 0;
const int BLOCK_SIZE = GASPI_GPU_BUFFERED;
const gaspi_cycles_t s0 = gaspi_get_cycles ();
while(size_left > 0)
{
int i;
for(i = 0; i < GASPI_CUDA_EVENTS; i++)
{
if(size_left > BLOCK_SIZE)
copy_size = BLOCK_SIZE;
else
copy_size = size_left;
if( cudaMemcpyAsync(host_ptr + gpu_offset, device_ptr + gpu_offset, copy_size, cudaMemcpyDeviceToHost, agpu->streams[queue]))
{
unlock_gaspi(&glb_gaspi_ctx.lockC[queue]);
return GASPI_ERROR;
}
glb_gaspi_ctx.ne_count_c[queue]++;
agpu->events[queue][i].segment_remote = segment_id_remote;
agpu->events[queue][i].segment_local = segment_id_local;
agpu->events[queue][i].size = copy_size;
agpu->events[queue][i].rank = rank;
agpu->events[queue][i].offset_local = offset_local+gpu_offset;
agpu->events[queue][i].offset_remote = offset_remote+gpu_offset;
agpu->events[queue][i].in_use =1;
cudaError_t err = cudaEventRecord(agpu->events[queue][i].event, agpu->streams[queue]);
if(err != cudaSuccess)
{
glb_gaspi_ctx.qp_state_vec[queue][rank] = GASPI_STATE_CORRUPT;
unlock_gaspi(&glb_gaspi_ctx.lockC[queue]);
return GASPI_ERROR;
}
gpu_offset += copy_size;
size_left -= copy_size;
if(size_left == 0)
break;
if(agpu->events[queue][i].ib_use)
{
struct ibv_wc wc;
int ne;
do
{
ne = ibv_poll_cq (glb_gaspi_ctx_ib.scqC[queue], 1, &wc);
glb_gaspi_ctx.ne_count_c[queue] -= ne;
if (ne == 0)
{
const gaspi_cycles_t s1 = gaspi_get_cycles ();
const gaspi_cycles_t tdelta = s1 - s0;
const float ms = (float) tdelta * glb_gaspi_ctx.cycles_to_msecs;
if (ms > timeout_ms)
{
unlock_gaspi (&glb_gaspi_ctx.lockC[queue]);
return GASPI_TIMEOUT;
}
}
} while(ne==0);
agpu->events[queue][i].ib_use = 0;
}
}
for(i = 0; i < GASPI_CUDA_EVENTS; i++)
{
cudaError_t error;
if ( agpu->events[queue][i].in_use == 1 )
{
do
{
error = cudaEventQuery(agpu->events[queue][i].event );
if( cudaSuccess == error )
{
if (_gaspi_event_send(&agpu->events[queue][i],queue))
{
unlock_gaspi (&glb_gaspi_ctx.lockC[queue]);
return GASPI_ERROR;
}
agpu->events[queue][i].in_use = 0;
}
else if(error == cudaErrorNotReady)
{
const gaspi_cycles_t s1 = gaspi_get_cycles ();
const gaspi_cycles_t tdelta = s1 - s0;
const float ms = (float) tdelta * glb_gaspi_ctx.cycles_to_msecs;
if (ms > timeout_ms)
{
unlock_gaspi (&glb_gaspi_ctx.lockC[queue]);
return GASPI_TIMEOUT;
}
}
else
{
unlock_gaspi (&glb_gaspi_ctx.lockC[queue]);
return GASPI_ERROR;
}
} while(error != cudaSuccess);
}
}
}
unlock_gaspi (&glb_gaspi_ctx.lockC[queue]);
return GASPI_SUCCESS;
}
