/**
* @brief Test the timer.
*
* An easy way to test the timer is to time the sleep function
* for a fixed amount of time. If the timer results a time that
* matches the sleep time within a reasonable error, the timer works.
*
* @return 0 on success, 1 on error.
*/
int trios_timer_test()
{
int rc = 0;
static const uint64_t sleep_us = 1000000; /* 1 second */
/* TODO: what should be the acceptable error (1/2 sec?) */
static const uint64_t error_ms = 500;
uint64_t error_ns = error_ms * 1000000;
uint64_t error_us = error_ms * 1000;
double error_sec = (double)error_ms/1000.0;
uint64_t start_ns=0, t_ns=0;
uint64_t start_us, t_us;
uint64_t start_ms, t_ms;
double start_sec, t_sec;
log_debug(timer_debug_level, "this is a timer test. expect a 1 second delay. starting...");
/* run to initialize timer */
start_ns = trios_get_time_ns();
if (usleep(sleep_us)) {
log_error(timer_debug_level, "Failed calling usleep()");
return 1;
}
/* run to get measurements */
start_ns = trios_get_time_ns();
start_us = trios_get_time_us();
start_ms = trios_get_time_ms();
start_sec = trios_get_time();
/* sleep one seconds */
usleep(sleep_us);
t_ns = trios_get_time_ns() - start_ns;
t_us = trios_get_time_us() - start_us;
t_ms = trios_get_time_ms() - start_ms;
t_sec = trios_get_time() - start_sec;
printf("slept for %llu seconds:\n", sleep_us);
printf("\tns = %llu\n", t_ns);
printf("\tus = %llu\n", t_us);
printf("\tms = %llu\n", t_ms);
printf("\tsec = %f\n", t_sec);
/* Make sure our values have a reasonable error */
if (labs(t_ns - (sleep_us * 1e3)) > error_ns) {
uint64_t actual_err = labs(t_ns - (sleep_us *1e3));
log_error(timer_debug_level,
"trios_timer failed ns timer test: err = %llu ns"
" > acceptable err = %llu",
actual_err, error_ns);
rc = 1;
}
if (labs(t_us - sleep_us) > error_us) {
uint64_t actual_err = labs(t_us - sleep_us);
log_error(timer_debug_level,
"trios_timer failed usec timer test: err = %llu usec"
" > acceptable err = %llu",
actual_err, error_us);
rc |= 1;
}
if (labs(t_ms - (sleep_us / 1e3)) > error_ms) {
uint64_t actual_err = labs(t_ms - (sleep_us / 1e3));
log_error(timer_debug_level,
"trios_timer failed ns timer test: err = %llu ms"
" > acceptable err = %llu",
actual_err, error_ms);
rc |= 1;
}
if (fabs(t_sec - ((double)sleep_us / 1e6)) > error_sec) {
double actual_err = fabs(t_sec - ((double)sleep_us/1e6));
log_error(timer_debug_level,
"trios_timer failed sec timer test: err = %g sec "
" > acceptable err = %g",
actual_err, error_sec);
rc |= 1;
}
log_debug(timer_debug_level, "timer test complete");
return rc;
}
void client(void) {
NNTI_result_t rc=NNTI_OK;
NNTI_status_t rdma_status;
NNTI_status_t send_status;
NNTI_status_t client_ack_status;
void *packed=NULL;
int32_t packed_size=0;
double op_timer;
// Teuchos::oblackholestream blackhole;
// std::ostream &out = ( rank == 1 ? std::cout : blackhole );
std::ostream &out = std::cout;
NNTI_connect(&trans_hdl, url, 5000, &server_hdl);
char *send_buf=(char *)malloc(NNTI_REQUEST_BUFFER_SIZE);
memset(send_buf, 0, NNTI_REQUEST_BUFFER_SIZE);
NNTI_register_memory(&trans_hdl, send_buf, NNTI_REQUEST_BUFFER_SIZE, 1, NNTI_SEND_SRC, NULL, &send_mr);
char *client_ack_buf=(char *)malloc(NNTI_REQUEST_BUFFER_SIZE);
memset(client_ack_buf, 0, NNTI_REQUEST_BUFFER_SIZE);
NNTI_register_memory(&trans_hdl, client_ack_buf, NNTI_REQUEST_BUFFER_SIZE, 1, NNTI_RECV_DST, NULL, &client_ack_mr);
char *get_dst_buf=(char *)malloc(get_size);
memset(get_dst_buf, 0, get_size);
NNTI_register_memory(&trans_hdl, get_dst_buf, get_size, 1, NNTI_GET_DST, NULL, &get_dst_mr);
char *put_src_buf=(char *)malloc(put_size);
memset(put_src_buf, 0, put_size);
NNTI_register_memory(&trans_hdl, put_src_buf, put_size, 1, NNTI_PUT_SRC, NULL, &put_src_mr);
/*
* Phase 1 - exchange buffer handles
*/
buffer_pack(&client_ack_mr, &packed, &packed_size, (xdrproc_t)&xdr_NNTI_buffer_t);
if (packed_size > NNTI_REQUEST_BUFFER_SIZE) {
log_error(nntiperf_debug_level, "buffer_pack() says encoded NNTI_buffer_t is larger than NNTI_REQUEST_BUFFER_SIZE");
MPI_Abort(MPI_COMM_WORLD, -10);
}
// send the server the recv_mr so it can send back it's ack_mr
memcpy(send_buf, packed, packed_size);
buffer_pack_free(packed, packed_size, (xdrproc_t)&xdr_NNTI_buffer_t);
rc=NNTI_send(&server_hdl, &send_mr, NULL);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_send() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
rc=NNTI_wait(&send_mr, NNTI_SEND_SRC, 5000, &send_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
// wait for the server to send back it's recv_mr
rc=NNTI_wait(&client_ack_mr, NNTI_RECV_DST, -1, &client_ack_status);
char *ptr=(char*)client_ack_status.start+client_ack_status.offset;
memcpy(&packed_size, ptr, sizeof(packed_size));
ptr += sizeof(packed_size);
memcpy(packed, ptr, packed_size);
ptr += packed_size;
buffer_unpack(packed, packed_size, &server_ack_mr, (xdrproc_t)&xdr_NNTI_buffer_t);
memcpy(&packed_size, ptr, sizeof(packed_size));
ptr += sizeof(packed_size);
memcpy(packed, ptr, packed_size);
ptr += packed_size;
buffer_unpack(packed, packed_size, &get_src_mr, (xdrproc_t)&xdr_NNTI_buffer_t);
memcpy(&packed_size, ptr, sizeof(packed_size));
ptr += sizeof(packed_size);
memcpy(packed, ptr, packed_size);
ptr += packed_size;
buffer_unpack(packed, packed_size, &put_dst_mr, (xdrproc_t)&xdr_NNTI_buffer_t);
// fprint_NNTI_buffer(logger_get_file(), "server_ack_mr",
// "received server ack hdl", &server_ack_mr);
// fprint_NNTI_buffer(logger_get_file(), "get_src_mr",
// "received get src hdl", &get_src_mr);
// fprint_NNTI_buffer(logger_get_file(), "put_dst_mr",
// "received put dst hdl", &put_dst_mr);
MPI_Barrier(MPI_COMM_WORLD);
/*
* Phase 2 - test sync request performance
*/
op_timer=trios_get_time();
for (int i=0;i<num_sends;i++) {
rc=NNTI_send(&server_hdl, &send_mr, NULL);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_send() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
rc=NNTI_wait(&send_mr, NNTI_SEND_SRC, 1000, &send_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time()-op_timer;
if (num_sends > 0) {
out << " sync requests per second == " << num_sends/op_timer << std::endl;
}
MPI_Barrier(MPI_COMM_WORLD);
/*
* Phase 3 - test async request performance
*/
op_timer=trios_get_time();
for (int i=0;i<num_sends;i++) {
rc=NNTI_send(&server_hdl, &send_mr, NULL);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_send() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
for (int i=0;i<num_sends;i++) {
rc=NNTI_wait(&send_mr, NNTI_SEND_SRC, 1000, &send_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time()-op_timer;
if (num_sends > 0) {
out << "async requests per second == " << num_sends/op_timer << std::endl;
}
MPI_Barrier(MPI_COMM_WORLD);
/*
* Phase 4 - test sync get performance
*/
// warm up the pipes
for (int i=0;i<num_gets;i++) {
rc=NNTI_get(&get_src_mr, client_rank*get_size, get_size, &get_dst_mr, 0);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_get() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
rc=NNTI_wait(&get_dst_mr, NNTI_GET_DST, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time();
for (int i=0;i<num_gets;i++) {
rc=NNTI_get(&get_src_mr, client_rank*get_size, get_size, &get_dst_mr, 0);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_get() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
rc=NNTI_wait(&get_dst_mr, NNTI_GET_DST, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time()-op_timer;
if (num_gets > 0) {
out << " sync get (" << get_size << " byte transfer) == " << (double)(num_gets*get_size)/one_mb/op_timer << " MBps" << std::endl;
}
MPI_Barrier(MPI_COMM_WORLD);
/*
* Phase 5 - test async get performance
*/
// warm up the pipes
for (int i=0;i<num_gets;i++) {
rc=NNTI_get(&get_src_mr, client_rank*get_size, get_size, &get_dst_mr, 0);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_get() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
for (int i=0;i<num_gets;i++) {
rc=NNTI_wait(&get_dst_mr, NNTI_GET_DST, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time();
for (int i=0;i<num_gets;i++) {
rc=NNTI_get(&get_src_mr, client_rank*get_size, get_size, &get_dst_mr, 0);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_get() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
for (int i=0;i<num_gets;i++) {
rc=NNTI_wait(&get_dst_mr, NNTI_GET_DST, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time()-op_timer;
if (num_gets > 0) {
out << "async get (" << get_size << " byte transfer) == " << (double)(num_gets*get_size)/one_mb/op_timer << " MBps" << std::endl;
}
MPI_Barrier(MPI_COMM_WORLD);
/*
* Phase 6 - test sync put performance
*/
// warm up the pipes
for (int i=0;i<num_puts;i++) {
rc=NNTI_put(&put_src_mr, 0, put_size, &put_dst_mr, client_rank*put_size);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_put() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
rc=NNTI_wait(&put_src_mr, NNTI_PUT_SRC, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time();
for (int i=0;i<num_puts;i++) {
rc=NNTI_put(&put_src_mr, 0, put_size, &put_dst_mr, client_rank*put_size);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_put() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
rc=NNTI_wait(&put_src_mr, NNTI_PUT_SRC, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time()-op_timer;
if (num_puts > 0) {
out << " sync put (" << put_size << " byte transfer) == " << (double)(num_puts*put_size)/one_mb/op_timer << " MBps" << std::endl;
}
MPI_Barrier(MPI_COMM_WORLD);
/*
* Phase 7 - test async put performance
*/
// warm up the pipes
for (int i=0;i<num_puts;i++) {
rc=NNTI_put(&put_src_mr, 0, put_size, &put_dst_mr, client_rank*put_size);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_put() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
for (int i=0;i<num_puts;i++) {
rc=NNTI_wait(&put_src_mr, NNTI_PUT_SRC, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time();
for (int i=0;i<num_puts;i++) {
rc=NNTI_put(&put_src_mr, 0, put_size, &put_dst_mr, client_rank*put_size);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_put() returned an error: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
for (int i=0;i<num_puts;i++) {
rc=NNTI_wait(&put_src_mr, NNTI_PUT_SRC, 1000, &rdma_status);
if (rc != NNTI_OK) {
log_error(nntiperf_debug_level, "NNTI_wait() did not return NNTI_OK: %d", rc);
MPI_Abort(MPI_COMM_WORLD, rc);
}
}
op_timer=trios_get_time()-op_timer;
if (num_puts > 0) {
out << "async put (" << put_size << " byte transfer) == " << (double)(num_puts*put_size)/one_mb/op_timer << " MBps" << std::endl;
}
MPI_Barrier(MPI_COMM_WORLD);
NNTI_unregister_memory(&send_mr);
free(send_buf);
NNTI_unregister_memory(&client_ack_mr);
free(client_ack_buf);
NNTI_unregister_memory(&get_dst_mr);
free(get_dst_buf);
NNTI_unregister_memory(&put_src_mr);
free(put_src_buf);
return;
}
