static int run(void)
{
int i, ret = 0;
ret = ft_init_fabric();
if (ret)
return ret;
ret = ft_exchange_keys(&remote);
if (ret)
goto out;
if (!(opts.options & FT_OPT_SIZE)) {
for (i = 0; i < TEST_CNT; i++) {
if (!ft_use_size(i, opts.sizes_enabled))
continue;
opts.transfer_size = test_size[i].size;
init_test(&opts, test_name, sizeof(test_name));
ret = run_test();
if (ret)
goto out;
}
} else {
init_test(&opts, test_name, sizeof(test_name));
ret = run_test();
if (ret)
goto out;
}
ft_sync();
ft_finalize();
out:
return ret;
}
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;
}
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;
}
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 send_recv()
{
struct fi_cq_entry comp;
int ret;
ret = fi_recv(ep, rx_buf, rx_size + ft_rx_prefix_size(),
mr_desc, 0, &rx_ctx);
if (ret)
return ret;
ft_sync();
fprintf(stdout, "Posting a send...\n");
ret = ft_post_tx(ep, remote_fi_addr, tx_size, NO_CQ_DATA, &tx_ctx);
if (ret)
return ret;
while ((tx_cq_cntr < tx_seq) || (rx_cq_cntr < rx_seq)) {
/* Wait for completion events on CQs */
ret = fi_wait(waitset, -1);
if (ret < 0) {
FT_PRINTERR("fi_wait", ret);
return ret;
}
/* Read the send completion entry */
ret = fi_cq_read(txcq, &comp, 1);
if (ret > 0) {
tx_cq_cntr++;
fprintf(stdout, "Received send completion event!\n");
} else if (ret < 0 && ret != -FI_EAGAIN) {
if (ret == -FI_EAVAIL) {
ret = ft_cq_readerr(txcq);
} else {
FT_PRINTERR("fi_cq_read", ret);
}
return ret;
}
/* Read the recv completion entry */
ret = fi_cq_read(rxcq, &comp, 1);
if (ret > 0) {
rx_cq_cntr++;
fprintf(stdout, "Received recv completion event!\n");
} else if (ret < 0 && ret != -FI_EAGAIN) {
if (ret == -FI_EAVAIL) {
ret = ft_cq_readerr(rxcq);
} else {
FT_PRINTERR("fi_cq_read", ret);
}
return ret;
}
}
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 = 0;
if (hints->ep_attr->type == FI_EP_MSG)
ret = ft_init_fabric_cm();
else
ret = ft_init_fabric();
if (ret)
return ret;
ret = overflow_test();
ft_sync();
ft_finalize();
return ret;
}
static int run(void)
{
int i, ret;
if (!opts.dst_addr) {
ret = ft_start_server();
if (ret)
return ret;
}
ret = opts.dst_addr ? ft_client_connect() : ft_server_connect();
if (ret)
return ret;
ret = ft_exchange_keys(&remote);
if (ret)
return ret;
if (!(opts.options & FT_OPT_SIZE)) {
for (i = 0; i < TEST_CNT; i++) {
if (!ft_use_size(i, opts.sizes_enabled))
continue;
opts.transfer_size = test_size[i].size;
init_test(&opts, test_name, sizeof(test_name));
ret = run_test();
if (ret)
goto out;
}
} else {
init_test(&opts, test_name, sizeof(test_name));
ret = run_test();
if (ret)
goto out;
}
ft_sync();
ft_finalize();
out:
fi_shutdown(ep, 0);
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;
}
static int av_removal_test(void)
{
int ret;
fprintf(stdout, "AV address removal: ");
hints = fi_dupinfo(base_hints);
if (!hints)
return -FI_ENOMEM;
ret = ft_init_fabric();
if (ret)
goto out;
if (opts.dst_addr) {
ret = ft_tx(ep, remote_fi_addr, opts.transfer_size, &tx_ctx);
if (ret) {
FT_PRINTERR("ft_tx", -ret);
goto out;
}
ret = fi_av_remove(av, &remote_fi_addr, 1, 0);
if (ret) {
FT_PRINTERR("fi_av_remove", ret);
goto out;
}
ret = ft_sync();
if (ret)
goto out;
ret = ft_init_av();
if (ret) {
FT_PRINTERR("ft_init_av", -ret);
goto out;
}
ret = ft_rx(ep, opts.transfer_size);
if (ret) {
FT_PRINTERR("ft_rx", -ret);
goto out;
}
} else {
ret = ft_rx(ep, opts.transfer_size);
if (ret) {
FT_PRINTERR("ft_rx", -ret);
goto out;
}
ret = fi_av_remove(av, &remote_fi_addr, 1, 0);
if (ret) {
FT_PRINTERR("fi_av_remove", ret);
goto out;
}
ret = ft_sync();
if (ret)
goto out;
ret = ft_init_av();
if (ret) {
FT_PRINTERR("ft_init_av", -ret);
goto out;
}
ret = ft_tx(ep, remote_fi_addr, opts.transfer_size, &tx_ctx);
if (ret) {
FT_PRINTERR("ft_tx", -ret);
goto out;
}
}
fprintf(stdout, "PASS\n");
(void) ft_sync();
out:
ft_free_res();
return ret;
}
static int run(void)
{
int ret;
ret = init_fabric();
if (ret)
return ret;
ret = ft_init_av();
if (ret)
return ret;
if (opts.dst_addr) {
printf("Searching for a bad msg\n");
ret = tagged_peek(0xbad);
if (ret != -FI_ENOMSG) {
FT_PRINTERR("FI_PEEK", ret);
return ret;
}
printf("Synchronizing with sender..\n");
ret = ft_sync();
if (ret)
return ret;
printf("Searching for a good msg\n");
ret = tagged_peek(0x900d);
if (ret != 1) {
FT_PRINTERR("FI_PEEK", ret);
return ret;
}
printf("Receiving msg\n");
ret = fi_trecv(ep, buf, rx_size, fi_mr_desc(mr), remote_fi_addr,
0x900d, 0, &rx_ctx);
if (ret) {
FT_PRINTERR("fi_trecv", ret);
return ret;
}
printf("Completing recv\n");
ret = ft_get_rx_comp(++rx_seq);
if (ret)
return ret;
} else {
printf("Sending tagged message\n");
ret = fi_tsend(ep, tx_buf, tx_size, fi_mr_desc(mr),
remote_fi_addr, 0x900d, &tx_ctx);
if (ret)
return ret;
printf("Synchronizing with receiver..\n");
ret = ft_sync();
if (ret)
return ret;
printf("Getting send completion\n");
ret = ft_get_tx_comp(tx_seq + 1);
if (ret)
return ret;
}
ft_finalize();
return 0;
}
static int run_test_loop(void)
{
int ret = 0;
uint64_t op_data = send_data ? opts.transfer_size : NO_CQ_DATA;
uint64_t op_tag = 0x1234;
char *op_buf;
int i, j;
for (i = 0; i < num_iters; i++) {
for (j = 0; j < concurrent_msgs; j++) {
op_buf = get_tx_buf(j);
if (ft_check_opts(FT_OPT_VERIFY_DATA))
ft_fill_buf(op_buf + ft_tx_prefix_size(),
opts.transfer_size);
ret = ft_post_tx_buf(ep, remote_fi_addr,
opts.transfer_size,
op_data, &tx_ctx_arr[j],
op_buf, mr_desc, op_tag);
if (ret) {
printf("ERROR send_msg returned %d\n", ret);
return ret;
}
}
ret = ft_sync();
if (ret)
return ret;
for (j = 0; j < concurrent_msgs; j++) {
op_buf = get_rx_buf(j);
ret = ft_post_rx_buf(ep, opts.transfer_size,
&rx_ctx_arr[j], op_buf,
mr_desc, op_tag);
if (ret) {
printf("ERROR recv_msg returned %d\n", ret);
return ret;
}
}
for (j = 0; j < concurrent_msgs; j++) {
ret = wait_recvs();
if (ret < 1)
return ret;
}
if (ft_check_opts(FT_OPT_VERIFY_DATA)) {
for (j = 0; j < concurrent_msgs; j++) {
op_buf = get_rx_buf(j);
if (ft_check_buf(op_buf + ft_rx_prefix_size(),
opts.transfer_size))
return -FI_EOTHER;
}
}
for (j = 0; j < concurrent_msgs; j++) {
ret = ft_get_tx_comp(tx_seq);
if (ret)
return ret;
}
if (i % 100 == 0)
printf("PID %d GOOD iter %d/%ld completed\n",
getpid(), i, num_iters);
}
(void) ft_sync();
printf("PID %d GOOD all done\n", getpid());
return ret;
}
