static int run_test()
{
int ret = 0, i;
if (opts.dst_addr) {
for (i = 0; i < ctx_cnt && !ret; i++) {
fprintf(stdout, "Posting send for ctx: %d\n", i);
ret = fi_send(tx_ep[i], buf, tx_size, fi_mr_desc(mr),
remote_rx_addr[i], NULL);
if (ret) {
FT_PRINTERR("fi_send", ret);
return ret;
}
ret = wait_for_comp(txcq_array[i]);
}
} else {
for (i = 0; i < ctx_cnt && !ret; i++) {
fprintf(stdout, "wait for recv completion for ctx: %d\n", i);
ret = wait_for_comp(rxcq_array[i]);
}
}
return ret;
}
static int run_test()
{
int ret = 0, i;
uint32_t data;
uint32_t *tb = (uint32_t *)tx_buf;
uint32_t *rb = (uint32_t *)rx_buf;
if (opts.dst_addr) {
for (i = 0; i < ctx_cnt && !ret; i++) {
fprintf(stdout, "Posting send for ctx: %d\n", i);
tb[0] = DATA + i;
ret = fi_send(tx_ep[i], tx_buf, tx_size, mr_desc,
remote_rx_addr[i], NULL);
if (ret) {
FT_PRINTERR("fi_send", ret);
return ret;
}
ret = wait_for_comp(txcq_array[i]);
}
} else {
for (i = 0; i < ctx_cnt && !ret; i++) {
fprintf(stdout, "wait for recv completion for ctx: %d\n", i);
ret = wait_for_comp(rxcq_array[i]);
data = DATA + i;
if (memcmp(&data, rx_buf, 4) != 0) {
fprintf(stdout, "failed compare expected 0x%x,"
" read 0x%x\n", data, rb[0]);
}
}
}
return ret;
}
static int init_av(void)
{
size_t addrlen;
int ret, i;
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(&sep->fid, tx_buf, &addrlen);
if (ret) {
FT_PRINTERR("fi_getname", ret);
return ret;
}
ret = fi_send(tx_ep[0], tx_buf, addrlen,
fi_mr_desc(mr), remote_fi_addr, NULL);
if (ret) {
FT_PRINTERR("fi_send", ret);
return ret;
}
ret = wait_for_comp(rxcq_array[0]);
if (ret)
return ret;
} else {
ret = wait_for_comp(rxcq_array[0]);
if (ret)
return ret;
ret = ft_av_insert(av, rx_buf, 1, &remote_fi_addr, 0, NULL);
if (ret)
return ret;
ret = fi_send(tx_ep[0], tx_buf, 1,
fi_mr_desc(mr), remote_fi_addr, NULL);
if (ret) {
FT_PRINTERR("fi_send", ret);
return ret;
}
}
for (i = 0; i < ctx_cnt; i++)
remote_rx_addr[i] = fi_rx_addr(remote_fi_addr, i, rx_ctx_bits);
ret = fi_recv(rx_ep[0], rx_buf, rx_size, fi_mr_desc(mr), 0, NULL);
if (ret) {
FT_PRINTERR("fi_recv", ret);
return ret;
}
ret = wait_for_comp(txcq_array[0]);
return ret;
}
double calc_bw(int rank, int num_pairs, int window_size, struct iovec *s_iov,
struct iovec *r_iov, int iov_cnt, char *s_buf, int s_buf_len,
char *r_buf, int r_buf_len, enum send_recv_type_e type)
{
uint64_t t_start = 0, t_end = 0, t = 0, maxtime = 0, *ts;
double bw = 0;
int c, i, j, target;
int loop, skip;
size_t cum_size = 0;
int mult = (DEFAULT_WINDOW / window_size) > 0 ? (DEFAULT_WINDOW /
window_size) : 1;
int __attribute__((unused)) fi_rc;
char r_fin[4] = {'b', 'b', 'b', 'b'};
char s_fin[4] = {'a', 'a', 'a', 'a'};
for (c = 0; c < iov_cnt; c++) {
for (i = 0; i < s_iov[c].iov_len; i++) {
((char *) s_iov[c].iov_base)[i] = 'a';
}
for (i = 0; i < r_iov[c].iov_len; i++) {
((char *) r_iov[c].iov_base)[i] = 'b';
}
/* Size will be all the receiver can hold */
cum_size += r_iov[c].iov_len;
}
for (c = 0; c < s_buf_len; c++) {
s_buf[c]= 'a';
}
for (c = 0; c < r_buf_len; c++) {
r_buf[c]= 'b';
}
if (cum_size > LARGE_THRESHOLD) {
loop = ITERS_LARGE * mult;
skip = WARMUP_ITERS_LARGE * mult;
} else {
loop = ITERS_SMALL * mult;
skip = WARMUP_ITERS_SMALL * mult;
}
ctpm_Barrier();
if (rank < num_pairs) {
target = rank + num_pairs;
for (i = 0; i < loop + skip; i++) {
if (i == skip) {
ctpm_Barrier();
t_start = get_time_usec();
}
for (j = 0; j < window_size; j++) {
switch (type) {
case SEND_RECV:
for (c = 0; c < iov_cnt; c++) {
fi_rc = fi_send(ep, s_iov[c].iov_base,
s_iov[c].iov_len, NULL,
fi_addrs[target],
NULL);
assert(!fi_rc);
}
break;
case SENDV_RECVV:
case SENDV_RECV:
fi_rc = fi_sendv(ep, s_iov, (void **) NULL,
iov_cnt, fi_addrs[target],
NULL);
assert(!fi_rc);
break;
case SEND_RECVV:
fi_rc = fi_send(ep, s_buf,
s_buf_len, NULL,
fi_addrs[target],
NULL);
assert(!fi_rc);
break;
default:
abort();
}
}
wait_for_comp(scq, type == SEND_RECV ? window_size * iov_cnt : window_size);
fi_rc = fi_recv(ep, r_fin, 4, NULL,
fi_addrs[target], NULL);
assert(!fi_rc);
wait_for_comp(rcq, 1);
}
t_end = get_time_usec();
t = t_end - t_start;
} else if (rank < num_pairs * 2) {
target = rank - num_pairs;
for (i = 0; i < loop + skip; i++) {
if (i == skip) {
ctpm_Barrier();
}
for (j = 0; j < window_size; j++) {
switch (type) {
case SEND_RECV:
for (c = 0; c < iov_cnt; c++) {
fi_rc = fi_recv(ep, r_iov[c].iov_base,
r_iov[c].iov_len, NULL,
fi_addrs[target],
NULL);
assert(!fi_rc);
}
break;
case SENDV_RECVV:
case SEND_RECVV:
fi_rc = fi_recvv(ep, r_iov, (void **) NULL, iov_cnt,
fi_addrs[target], NULL);
assert(!fi_rc);
break;
case SENDV_RECV:
fi_rc = fi_recv(ep, r_buf,
r_buf_len, NULL,
fi_addrs[target],
NULL);
assert(!fi_rc);
break;
default:
abort();
}
}
wait_for_comp(rcq, type == 0 ? window_size * iov_cnt : window_size);
fi_rc = fi_send(ep, s_fin, 4, NULL,
fi_addrs[target], NULL);
assert(!fi_rc);
wait_for_comp(scq, 1);
}
} else {
ctpm_Barrier();
}
ts = malloc(sizeof(t) * numprocs);
assert(ts);
ctpm_Allgather(&t, sizeof(t), ts);
if (!myid) {
for (i = 0; i < numprocs; i++) {
if (ts[i] > maxtime) {
maxtime = ts[i];
}
}
}
free(ts);
if (rank == 0) {
double tmp = num_pairs * cum_size / 1e6;
tmp = tmp * loop * window_size;
bw = tmp / (maxtime / 1e6);
return bw;
}
return 0;
}
