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;
}
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 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 ft_sync()
{
int ret;
if (opts.dst_addr) {
ret = ft_tx(1);
if (ret)
return ret;
ret = ft_rx(1);
} else {
ret = ft_rx(1);
if (ret)
return ret;
ret = ft_tx(1);
}
return ret;
}
int ft_sync()
{
int ret;
if (opts.dst_addr) {
ret = ft_tx(ep, remote_fi_addr, 1, &tx_ctx);
if (ret)
return ret;
ret = ft_rx(ep, 1);
} else {
ret = ft_rx(ep, 1);
if (ret)
return ret;
ret = ft_tx(ep, remote_fi_addr, 1, &tx_ctx);
}
return ret;
}
/* TODO: retry send for unreliable endpoints */
int ft_init_av(void)
{
size_t addrlen;
int ret;
if (opts.dst_addr) {
ret = ft_av_insert(av, fi->dest_addr, 1, &remote_fi_addr, 0, NULL);
if (ret)
return ret;
addrlen = FT_MAX_CTRL_MSG;
ret = fi_getname(&ep->fid, (char *) tx_buf + ft_tx_prefix_size(),
&addrlen);
if (ret) {
FT_PRINTERR("fi_getname", ret);
return ret;
}
ret = (int) ft_tx(addrlen);
if (ret)
return ret;
ret = ft_rx(1);
} else {
ret = (int) ft_rx(FT_MAX_CTRL_MSG);
if (ret)
return ret;
ret = ft_av_insert(av, (char *) rx_buf + ft_rx_prefix_size(),
1, &remote_fi_addr, 0, NULL);
if (ret)
return ret;
ret = (int) ft_tx(1);
}
return ret;
}
ssize_t ft_rma(enum ft_rma_opcodes op, struct fid_ep *ep, size_t size,
struct fi_rma_iov *remote, void *context)
{
int ret;
ret = ft_post_rma(op, ep, size, remote, context);
if (ret)
return ret;
if (op == FT_RMA_WRITEDATA) {
ret = ft_rx(ep, 0);
if (ret)
return ret;
}
ret = ft_get_tx_comp(tx_seq);
if (ret)
return ret;
return 0;
}
static int init_av(void)
{
int ret;
int i;
if (opts.dst_addr) {
ret = ft_av_insert(av, fi->dest_addr, 1, &addr_array[0], 0, NULL);
if (ret)
return ret;
}
for (i = 0; i < ep_cnt; i++) {
addrlen = tx_size;
ret = fi_getname(&ep_array[i]->fid, tx_buf + ft_tx_prefix_size(),
&addrlen);
if (ret) {
FT_PRINTERR("fi_getname", ret);
return ret;
}
if (opts.dst_addr) {
ret = ft_tx(ep_array[0], addr_array[0], addrlen, &tx_ctx);
if (ret)
return ret;
if (rx_shared_ctx)
ret = ft_rx(srx_ctx, rx_size);
else
ret = ft_rx(ep_array[0], rx_size);
if (ret)
return ret;
/* Skip the first address since we already have it in AV */
if (i) {
ret = ft_av_insert(av, rx_buf + ft_rx_prefix_size(), 1,
&addr_array[i], 0, NULL);
if (ret)
return ret;
}
} else {
if (rx_shared_ctx)
ret = ft_rx(srx_ctx, rx_size);
else
ret = ft_rx(ep_array[0], rx_size);
if (ret)
return ret;
ret = ft_av_insert(av, rx_buf + ft_rx_prefix_size(), 1,
&addr_array[i], 0, NULL);
if (ret)
return ret;
ret = ft_tx(ep_array[0], addr_array[0], addrlen, &tx_ctx);
if (ret)
return ret;
}
}
/* ACK */
if (opts.dst_addr) {
ret = ft_tx(ep_array[0], addr_array[0], 1, &tx_ctx);
} else {
if (rx_shared_ctx)
ret = ft_rx(srx_ctx, rx_size);
else
ret = ft_rx(ep_array[0], rx_size);
}
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;
}
