mca_btl_base_descriptor_t* mca_btl_mx_prepare_dst( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
struct mca_mpool_base_registration_t* registration,
struct ompi_convertor_t* convertor,
uint8_t order,
size_t reserve,
size_t* size,
uint32_t flags)
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*)btl;
mca_btl_mx_frag_t* frag;
mx_return_t mx_return;
mx_segment_t mx_segment;
int rc;
MCA_BTL_MX_FRAG_ALLOC_USER(btl, frag, rc);
if( OPAL_UNLIKELY(NULL == frag) ) {
return NULL;
}
frag->segment[0].seg_len = *size;
ompi_convertor_get_current_pointer( convertor, (void**)&(frag->segment[0].seg_addr.pval) );
frag->segment[0].seg_key.key64 = (uint64_t)(intptr_t)frag;
mx_segment.segment_ptr = frag->segment[0].seg_addr.pval;
mx_segment.segment_length = frag->segment[0].seg_len;
mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1,
frag->segment[0].seg_key.key64,
BTL_MX_PUT_MASK, NULL, &(frag->mx_request) );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "Fail to re-register a fragment with the MX NIC ...\n" );
MCA_BTL_MX_FRAG_RETURN( btl, frag );
return NULL;
}
#ifdef HAVE_MX_FORGET
{
mx_return = mx_forget( mx_btl->mx_endpoint, &(frag->mx_request) );
if( OPAL_UNLIKELY(MX_SUCCESS != mx_return) ) {
opal_output( 0, "mx_forget failed in mca_btl_mx_prepare_dst with error %d (%s)\n",
mx_return, mx_strerror(mx_return) );
return NULL;
}
}
#endif
/* Allow the fragment to be recycled using the mca_btl_mx_free function */
frag->type = MCA_BTL_MX_SEND;
frag->base.des_dst = frag->segment;
frag->base.des_dst_cnt = 1;
frag->base.des_flags = flags;
frag->base.order = MCA_BTL_NO_ORDER;
return &frag->base;
}
static inline void
sender(mx_endpoint_t ep,mx_endpoint_addr_t dest, int iter, int len,
int blocking, int bothways, uint32_t match_val)
{
int count, cur_req, num_req;
mx_status_t stat;
mx_request_t req[NUM_SREQ];
mx_segment_t seg;
uint32_t result;
struct metadata info;
char *buffer;
uint32_t usec;
double bw, pkts_per_sec;
#if MX_THREAD_SAFE
MX_THREAD_T thread;
struct mx_thread_arg args;
#endif
struct bwinfo *bwinfo = NULL;
buffer = malloc(len * NUM_SREQ);
if (buffer == NULL) {
fprintf(stderr, "Can't allocate buffers\n");
exit(1);
}
seg.segment_ptr = &info;
seg.segment_length = sizeof(info);
info.len = htonl(len);
info.iter = htonl(iter);
info.verify = htonl(Verify);
info.bothways = htonl(bothways);
if (bothways) {
#if MX_THREAD_SAFE
args.ep = ep;
args.dest = dest;
args.iter = iter;
args.len = len;
args.blocking = blocking;
MX_THREAD_CREATE(&thread, &start_recv_thread, &args);
#else
fprintf(stderr, "bothways not supported\n");
exit(1);
#endif
}
mx_isend(ep, &seg, 1, dest, match_val, NULL, &req[0]);
/* wait for the send to complete */
mx_test_or_wait(blocking, ep, &req[0], MX_INFINITE, &stat, &result);
if (!result) {
fprintf(stderr, "mx_wait failed\n");
exit(1);
}
if (stat.code != MX_STATUS_SUCCESS) {
fprintf(stderr, "isend failed with status %s\n", mx_strstatus(stat.code));
exit(1);
}
MX_MUTEX_LOCK(&stream_mutex);
++threads_running;
MX_MUTEX_UNLOCK(&stream_mutex);
while(threads_running != num_threads)
/* spin */;
#if DO_HANDSHAKE
/* wait for the receiver to get ready */
seg.segment_ptr = &info;
seg.segment_length = sizeof(info);
mx_irecv(ep, &seg, 1, match_val, MX_MATCH_MASK_NONE, 0, &req[0]);
/* wait for the receive to complete */
mx_test_or_wait(blocking, ep, &req[0], MX_INFINITE, &stat, &result);
if (!result) {
fprintf(stderr, "mx_wait failed\n");
exit(1);
}
if (stat.code != MX_STATUS_SUCCESS) {
fprintf(stderr, "irecv failed with status %s\n", mx_strstatus(stat.code));
exit(1);
}
#endif
num_req = NUM_SREQ;
if (num_req > iter)
num_req = iter;
for (cur_req = 0; cur_req < num_req; cur_req++) {
seg.segment_ptr = &buffer[cur_req * len];
seg.segment_length = len;
if (Verify)
mx_fill_buffer(seg.segment_ptr, len);
mx_isend(ep, &seg, 1, dest, match_val, NULL,
&req[cur_req]);
}
for (count = 0; count < iter; count++) {
/* wait for the send to complete */
cur_req = count & (NUM_SREQ - 1);
mx_test_or_wait(blocking, ep, &req[cur_req],
MX_INFINITE, &stat, &result);
if (!result) {
fprintf(stderr, "mx_wait failed\n");
exit(1);
}
if (stat.code != MX_STATUS_SUCCESS) {
fprintf(stderr, "isend failed with status %s\n", mx_strstatus(stat.code));
exit(1);
}
/* hack since mx_cancel does not work */
if ((count + NUM_SREQ) >= iter)
continue;
seg.segment_ptr = &buffer[cur_req * len];
seg.segment_length = len;
if (Verify)
mx_fill_buffer(seg.segment_ptr, len);
mx_isend(ep, &seg, 1, dest, match_val, NULL, &req[cur_req]);
}
seg.segment_ptr = &info;
seg.segment_length = sizeof(info);
mx_irecv(ep, &seg, 1, match_val, MX_MATCH_MASK_NONE, 0, &req[0]);
/* wait for the receive to complete */
mx_test_or_wait(blocking, ep, &req[0], MX_INFINITE, &stat, &result);
if (!result) {
fprintf(stderr, "mx_wait failed\n");
exit(1);
}
if (stat.code != MX_STATUS_SUCCESS) {
fprintf(stderr, "irecv failed with status %s\n", mx_strstatus(stat.code));
exit(1);
}
usec = ntohl(info.usec);
bw = ((double)iter * (double)len) / (double) usec;
pkts_per_sec = iter / ((double) usec / 1000000.0);
if (match_val == MATCH_VAL_THREAD)
return;
if (match_val == MATCH_VAL_MAIN && bothways)
#if MX_THREAD_SAFE
if(bothways) {
printf("Send: %8d %5.3f %5.3f\n",
len, bw, pkts_per_sec);
MX_THREAD_JOIN(thread);
bwinfo = &global_bwinfo;
}
#endif
if (bwinfo) {
printf("Recv: %8d %5.3f %5.3f\n",
len, bwinfo->bandwidth, bwinfo->pkts_per_sec);
bw += bwinfo->bandwidth;
pkts_per_sec += bwinfo->pkts_per_sec;
}
printf("Total: %8d %5.3f %5.3f\n", len, bw, pkts_per_sec);
}
int mca_btl_mx_register( struct mca_btl_base_module_t* btl,
mca_btl_base_tag_t tag,
mca_btl_base_module_recv_cb_fn_t cbfunc,
void* cbdata )
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*) btl;
#if 0
if( (NULL != cbfunc) && ( 0 == mca_btl_mx_component.mx_use_unexpected) ) {
#endif
if( NULL != cbfunc ) {
mca_btl_mx_frag_t* frag;
mx_return_t mx_return;
mx_segment_t mx_segment;
int i, rc;
/* Post the receives if there is no unexpected handler */
for( i = 0; i < mca_btl_mx_component.mx_max_posted_recv; i++ ) {
MCA_BTL_MX_FRAG_ALLOC_EAGER( mx_btl, frag, rc );
if( NULL == frag ) {
opal_output( 0, "mca_btl_mx_register: unable to allocate more eager fragments\n" );
if( 0 == i ) {
return OMPI_ERROR;
}
break; /* some fragments are already registered. Try to continue... */
}
frag->base.des_dst = frag->segment;
frag->base.des_dst_cnt = 1;
frag->base.des_src = NULL;
frag->base.des_src_cnt = 0;
frag->mx_frag_list = NULL;
frag->type = MCA_BTL_MX_RECV;
mx_segment.segment_ptr = (void*)(frag+1);
mx_segment.segment_length = mx_btl->super.btl_eager_limit;
mx_return = mx_irecv( mx_btl->mx_endpoint, &mx_segment, 1,
0x01ULL, BTL_MX_RECV_MASK,
frag, &(frag->mx_request) );
if( MX_SUCCESS != mx_return ) {
opal_output( 0, "mca_btl_mx_register: mx_irecv failed with status %d (%s)\n",
mx_return, mx_strerror(mx_return) );
MCA_BTL_MX_FRAG_RETURN( mx_btl, frag );
return OMPI_ERROR;
}
}
}
return OMPI_SUCCESS;
}
/**
* Allocate a segment.
*
* @param btl (IN) BTL module
* @param size (IN) Request segment size.
*/
mca_btl_base_descriptor_t* mca_btl_mx_alloc( struct mca_btl_base_module_t* btl,
struct mca_btl_base_endpoint_t* endpoint,
uint8_t order,
size_t size,
uint32_t flags)
{
mca_btl_mx_module_t* mx_btl = (mca_btl_mx_module_t*) btl;
mca_btl_mx_frag_t* frag;
int rc;
MCA_BTL_MX_FRAG_ALLOC_EAGER(mx_btl, frag, rc);
if( OPAL_UNLIKELY(NULL == frag) ) {
return NULL;
}
frag->segment[0].seg_len =
size <= mx_btl->super.btl_eager_limit ?
size : mx_btl->super.btl_eager_limit ;
frag->segment[0].seg_addr.pval = (void*)(frag+1);
frag->base.des_src = frag->segment;
frag->base.des_src_cnt = 1;
frag->base.des_flags = flags;
frag->base.order = MCA_BTL_NO_ORDER;
return (mca_btl_base_descriptor_t*)frag;
}
static inline void
receiver(mx_endpoint_t ep, int blocking, uint32_t match_val, uint32_t filter)
{
int count, len, iter, cur_req, num_req;
mx_status_t stat;
mx_request_t req[NUM_RREQ];
mx_request_t sreq;
mx_segment_t seg;
uint32_t result, usec;
struct timeval start_time, end_time;
double bw, pkts_per_sec;
char *buffer;
struct metadata info;
int bothways;
#if MX_THREAD_SAFE
struct mx_thread_arg args;
MX_THREAD_T thread;
#endif
uint64_t nic;
uint32_t eid;
seg.segment_ptr = &info;
seg.segment_length = sizeof(info);
mx_irecv(ep, &seg, 1, match_val, MX_MATCH_MASK_NONE, 0, &req[0]);
/* wait for the receive to complete */
mx_test_or_wait(blocking, ep, &req[0], MX_INFINITE, &stat, &result);
if (!result) {
fprintf(stderr, "mx_wait failed\n");
exit(1);
}
if (stat.code != MX_STATUS_SUCCESS) {
fprintf(stderr, "irecv failed with status %s\n", mx_strstatus(stat.code));
exit(1);
}
if (filter != ~0) {
/* filter == ~0 means recv threads on master */
mx_decompose_endpoint_addr(stat.source, &nic, &eid);
mx_connect(ep, nic, eid, filter, MX_INFINITE, &stat.source);
}
len = ntohl(info.len);
iter = ntohl(info.iter);
Verify = ntohl(info.verify);
bothways = ntohl(info.bothways);
if (do_verbose)
printf("Starting test: len = %d, iter = %d\n", len, iter);
if (do_verbose && Verify) {
printf("Verifying results\n");
}
buffer = malloc(len * NUM_RREQ);
if (buffer == NULL) {
fprintf(stderr, "Can't allocate buffers\n");
exit(1);
}
if (bothways) {
#if MX_THREAD_SAFE
args.ep = ep;
args.dest = stat.source;
args.iter = iter;
args.len = len;
args.blocking = blocking;
num_threads++;
MX_THREAD_CREATE(&thread, &start_send_thread, &args);
#else
fprintf(stderr,"bothways not supported\n");
exit(1);
#endif
}
/* pre-post our receives */
num_req = NUM_RREQ;
if (num_req > iter)
num_req = iter;
for (cur_req = 0; cur_req < num_req; cur_req++) {
seg.segment_ptr = &buffer[cur_req * len];
seg.segment_length = len;
mx_irecv(ep, &seg, 1, match_val, MX_MATCH_MASK_NONE, 0,
&req[cur_req]);
}
MX_MUTEX_LOCK(&stream_mutex);
++threads_running;
MX_MUTEX_UNLOCK(&stream_mutex);
while(threads_running != num_threads)
/* spin */;
#if DO_HANDSHAKE
/* post a send to let the sender know we are ready */
seg.segment_ptr = &info;
seg.segment_length = sizeof(info);
sreq = 0;
mx_isend(ep, &seg, 1, stat.source, match_val, NULL, &sreq);
mx_test_or_wait(blocking, ep, &sreq, MX_INFINITE, &stat, &result);
if (!result) {
fprintf(stderr, "mx_wait failed\n");
exit(1);
}
if (stat.code != MX_STATUS_SUCCESS) {
fprintf(stderr, "isend failed with status %s\n", mx_strstatus(stat.code));
exit(1);
}
#endif
/* start the test */
gettimeofday(&start_time, NULL);
for (count = 0; count < iter; count++) {
/* wait for the receive to complete */
cur_req = count & (NUM_RREQ - 1);
mx_test_or_wait(blocking, ep, &req[cur_req],
MX_INFINITE, &stat, &result);
if (!result) {
fprintf(stderr, "mx_wait failed\n");
exit(1);
}
if (stat.code != MX_STATUS_SUCCESS) {
fprintf(stderr, "irecv failed with status %s\n", mx_strstatus(stat.code));
exit(1);
}
if (stat.xfer_length != len) {
fprintf(stderr, "bad len %d != %d\n", stat.xfer_length, len);
exit(1);
}
/* hack since mx_cancel does not work */
if ((count + NUM_RREQ) > iter)
continue;
seg.segment_ptr = &buffer[cur_req * len];
seg.segment_length = len;
if (Verify)
mx_check_buffer(seg.segment_ptr, len);
mx_irecv(ep, &seg, 1, match_val, MX_MATCH_MASK_NONE, 0,
&req[cur_req]);
}
gettimeofday(&end_time, NULL);
usec = end_time.tv_usec - start_time.tv_usec;
usec += (end_time.tv_sec - start_time.tv_sec) * 1000000;
bw = ((double)iter * (double)len) / (double) usec;
pkts_per_sec = iter / ((double) usec / 1000000.0);
global_bwinfo.bandwidth = bw;
global_bwinfo.pkts_per_sec = pkts_per_sec;
/* printf("%8d %5.3f %5.3f\n", len, bw, pkts_per_sec);*/
#if 0 /* mx_cancel assert(0)'s */
for (cur_req = 0; cur_req < num_req; cur_req++) {
mx_cancel(ep, &req[cur_req]);
}
#endif
info.usec = htonl(usec);
seg.segment_ptr = &info;
seg.segment_length = sizeof(info);
sreq = 0;
mx_isend(ep, &seg, 1, stat.source, match_val, NULL, &sreq);
mx_test_or_wait(blocking, ep, &sreq, MX_INFINITE, &stat, &result);
free(buffer);
#if MX_THREAD_SAFE
if(bothways)
MX_THREAD_JOIN(thread);
#endif
}
