static int run_test(void)
{
int ret, i;
ret = sync_test();
if (ret) {
fprintf(stderr, "sync_test failed!\n");
goto out;
}
ft_start();
if (opts.dst_addr) {
for (i = 0; i < opts.iterations; i++) {
ret = ft_tx(ep, remote_fi_addr, opts.transfer_size, &tx_ctx);
if (ret)
goto out;
}
} else {
ret = wait_for_recv_completion(opts.iterations);
if (ret)
goto out;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations,
&start, &end, 1, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations,
&start, &end, 1);
out:
return ret;
}
static int run_test(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
ft_start();
for (i = 0; i < opts.iterations; i++) {
ret = ft_rma(opts.rma_op, ep, opts.transfer_size, &remote, ep);
if (ret)
return ret;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
1, opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations,
&start, &end, 1);
return 0;
}
static int run_test(void)
{
int ret, i;
ret = sync_test();
if (ret)
return ret;
clock_gettime(CLOCK_MONOTONIC, &start);
for (i = 0; i < opts.iterations; i++) {
ret = opts.dst_addr ? send_xfer(opts.transfer_size) :
recv_xfer(opts.transfer_size);
if (ret)
return ret;
ret = opts.dst_addr ? recv_xfer(opts.transfer_size) :
send_xfer(opts.transfer_size);
if (ret)
return ret;
}
clock_gettime(CLOCK_MONOTONIC, &end);
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 2, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations, &start, &end, 2);
return 0;
}
int pingpong(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
ft_start();
for (i = 0; i < opts.iterations; i++) {
ret = opts.dst_addr ?
ft_tx(opts.transfer_size) : ft_rx(opts.transfer_size);
if (ret)
return ret;
ret = opts.dst_addr ?
ft_rx(opts.transfer_size) : ft_tx(opts.transfer_size);
if (ret)
return ret;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 2, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations, &start, &end, 2);
return 0;
}
static int stream(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
ft_start();
for (i = 0; i < opts.iterations; i++) {
ret = opts.dst_addr ?
ft_tx(ep, remote_fi_addr, opts.transfer_size, &tx_ctx) : ft_rx(ep, opts.transfer_size);
if (ret)
return ret;
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 1,
opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations, &start, &end, 1);
return 0;
}
int pingpong(void)
{
int ret, i;
ret = ft_sync();
if (ret)
return ret;
if (opts.dst_addr) {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
if (opts.transfer_size < fi->tx_attr->inject_size)
ret = ft_inject(opts.transfer_size);
else
ret = ft_tx(opts.transfer_size);
if (ret)
return ret;
ret = ft_rx(opts.transfer_size);
if (ret)
return ret;
}
} else {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
ret = ft_rx(opts.transfer_size);
if (ret)
return ret;
if (opts.transfer_size < fi->tx_attr->inject_size)
ret = ft_inject(opts.transfer_size);
else
ret = ft_tx(opts.transfer_size);
if (ret)
return ret;
}
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end, 2, opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations, &start, &end, 2);
return 0;
}
static int run_test(void)
{
int ret, i;
ret = sync_test();
if (ret)
return ret;
clock_gettime(CLOCK_MONOTONIC, &start);
for (i = 0; i < opts.iterations; i++) {
switch (op_type) {
case FT_RMA_WRITE:
ret = write_data(opts.transfer_size);
break;
case FT_RMA_WRITEDATA:
ret = write_data_with_cq_data(opts.transfer_size);
if (ret)
return ret;
ret = wait_remote_writedata_completion();
break;
case FT_RMA_READ:
ret = read_data(opts.transfer_size);
break;
}
if (ret)
return ret;
ret = ft_wait_for_comp(txcq, 1);
if (ret)
return ret;
}
clock_gettime(CLOCK_MONOTONIC, &end);
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
1, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations,
&start, &end, 1);
return 0;
}
static int run_op(void)
{
int ret, i;
count = (size_t *) malloc(sizeof(size_t));
sync_test();
clock_gettime(CLOCK_MONOTONIC, &start);
switch (op_type) {
case FI_MIN:
case FI_MAX:
case FI_ATOMIC_READ:
case FI_ATOMIC_WRITE:
ret = is_valid_base_atomic_op(op_type);
if (ret > 0) {
for (i = 0; i < opts.iterations; i++) {
ret = execute_base_atomic_op(op_type);
if (ret)
break;
}
}
ret = is_valid_fetch_atomic_op(op_type);
if (ret > 0) {
for (i = 0; i < opts.iterations; i++) {
ret = execute_fetch_atomic_op(op_type);
if (ret)
break;
}
}
break;
case FI_CSWAP:
ret = is_valid_compare_atomic_op(op_type);
if (ret > 0) {
for (i = 0; i < opts.iterations; i++) {
ret = execute_compare_atomic_op(op_type);
if(ret)
break;
}
}
break;
default:
ret = -EINVAL;
goto out;
}
clock_gettime(CLOCK_MONOTONIC, &end);
if (ret)
goto out;
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
op_type == FI_CSWAP ? 1 : 2, opts.argc, opts.argv);
else
show_perf(test_name, opts.transfer_size, opts.iterations,
&start, &end, op_type == FI_CSWAP ? 1 : 2);
ret = 0;
out:
free(count);
return ret;
}
int bandwidth(void)
{
int ret, i, j;
ret = ft_sync();
if (ret)
return ret;
/* The loop structured allows for the possibility that the sender
* immediately overruns the receiving side on the first transfer (or
* the entire window). This could result in exercising parts of the
* provider's implementation of FI_RM_ENABLED. For better or worse,
* some MPI-level benchmarks tend to use this type of loop for measuring
* bandwidth. */
if (opts.dst_addr) {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
for(j = 0; j < opts.window_size; j++) {
if (opts.transfer_size < fi->tx_attr->inject_size)
ret = ft_inject(opts.transfer_size);
else
ret = ft_post_tx(opts.transfer_size);
if (ret)
return ret;
}
ret = ft_get_tx_comp(tx_seq);
if (ret)
return ret;
ret = ft_rx(4);
if (ret)
return ret;
}
} else {
for (i = 0; i < opts.iterations + opts.warmup_iterations; i++) {
if (i == opts.warmup_iterations)
ft_start();
for(j = 0; j < opts.window_size; j++) {
ret = ft_post_rx(opts.transfer_size);
if (ret)
return ret;
}
ret = ft_get_rx_comp(rx_seq-1); /* rx_seq is always one ahead */
if (ret)
return ret;
ret = ft_tx(4);
if (ret)
return ret;
}
}
ft_stop();
if (opts.machr)
show_perf_mr(opts.transfer_size, opts.iterations, &start, &end,
opts.window_size, opts.argc, opts.argv);
else
show_perf(NULL, opts.transfer_size, opts.iterations, &start, &end,
opts.window_size);
return 0;
}
