MPI_Request ref = (MPI_Request)a;
MPI_Request req = (MPI_Request)b;
XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
MPI_Request ref = (MPI_Request)a;
MPI_Request req = (MPI_Request)b;
XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
MPI_Request ref = (MPI_Request)a;
MPI_Request req = (MPI_Request)b;
XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
MPI_Request ref = (MPI_Request)a;
MPI_Request req = (MPI_Request)b;
XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src,req->src, ref->tag, req->tag);
+ // Iterate over all the sections that were specified and find the right
+ // value. (fact.factor represents the interval sizes; we want to find the
+ // section that has fact.factor <= size and no other such fact.factor <= size)
+ // Note: parse_factor() (used before) already sorts the dynar we iterate over!
XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
return current;
}else{
XBT_DEBUG("os : %f <= %ld return %f", size, fact.factor, current);
return current;
}else{
- current=fact.values[0]+fact.values[1]*size;
+ // If the next section is too large, the current section must be used.
+ // Hence, save the cost, as we might have to use it.
+ current = fact.values[0]+fact.values[1]*size;
}
}
XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
}
}
XBT_DEBUG("os : %f > %ld return %f", size, fact.factor, current);
+ // Iterate over all the sections that were specified and find the right
+ // value. (fact.factor represents the interval sizes; we want to find the
+ // section that has fact.factor <= size and no other such fact.factor <= size)
+ // Note: parse_factor() (used before) already sorts the dynar we iterate over!
XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
return current;
}else{
XBT_DEBUG("ois : %f <= %ld return %f", size, fact.factor, current);
return current;
}else{
- current=fact.values[0]+fact.values[1]*size;
+ // If the next section is too large, the current section must be used.
+ // Hence, save the cost, as we might have to use it.
+ current = fact.values[0]+fact.values[1]*size;
}
}
XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
}
}
XBT_DEBUG("ois : %f > %ld return %f", size, fact.factor, current);
+ // Iterate over all the sections that were specified and find the right
+ // value. (fact.factor represents the interval sizes; we want to find the
+ // section that has fact.factor <= size and no other such fact.factor <= size)
+ // Note: parse_factor() (used before) already sorts the dynar we iterate over!
XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
return current;
XBT_DEBUG("or : %f <= %ld return %f", size, fact.factor, current);
return current;
}
XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
}
XBT_DEBUG("or : %f > %ld return %f", size, fact.factor, current);
if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
smpi_bench_end();
time = SIMIX_get_clock();
if (smpi_process_initialized() && !smpi_process_finalized() && !smpi_process_get_sampling()) {
smpi_bench_end();
time = SIMIX_get_clock();
- //to avoid deadlocks if called too many times
+ // to avoid deadlocks if used as a break condition, such as
+ // while (MPI_Wtime(...) < time_limit) {
+ // ....
+ // }
+ // because the time will not normally advance when only calls to MPI_Wtime
+ // are made -> deadlock (MPI_Wtime never reaches the time limit)
if (request->flags & RECV) {
print_request("New recv", request);
if (request->flags & RECV) {
print_request("New recv", request);
if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
//We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
mailbox = smpi_process_mailbox_small();
if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")){
//We have to check both mailboxes (because SSEND messages are sent to the large mbox). begin with the more appropriate one : the small one.
mailbox = smpi_process_mailbox_small();
- XBT_DEBUG("Is there a corresponding send already posted the small mailbox %p (in case of SSEND)?", mailbox);
- smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
+ XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
+ smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
if(action ==NULL){
mailbox = smpi_process_mailbox();
XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
if(action ==NULL){
if(action ==NULL){
mailbox = smpi_process_mailbox();
XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
if(action ==NULL){
}else{
mailbox = smpi_process_mailbox_small();
XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
}else{
mailbox = smpi_process_mailbox_small();
XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
- smx_action_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
+ smx_synchro_t action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
+ //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
+ double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
+ if(sleeptime!=0.0){
+ simcall_process_sleep(sleeptime);
+ XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
+ }
+
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
request->real_size=request->size;
smpi_datatype_use(request->old_type);
smpi_comm_use(request->comm);
request->action = simcall_comm_irecv(mailbox, request->buf,
&request->real_size, &match_recv,
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
request->real_size=request->size;
smpi_datatype_use(request->old_type);
smpi_comm_use(request->comm);
request->action = simcall_comm_irecv(mailbox, request->buf,
&request->real_size, &match_recv,
- //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
- double sleeptime = request->detached ? smpi_or(request->size) : 0.0;
- if(sleeptime!=0.0){
- simcall_process_sleep(sleeptime);
- XBT_DEBUG("receiving size of %zu : sleep %f ", request->size, smpi_or(request->size));
- }
- #ifdef HAVE_TRACING
- int rank = request->src;
- if (TRACE_smpi_view_internals()) {
- TRACE_smpi_send(rank, rank, receiver,request->size);
- }
- #endif
+ int rank = request->src;
+ if (TRACE_smpi_view_internals()) {
+ TRACE_smpi_send(rank, rank, receiver,request->size);
+ }
+
+ //if we are giving back the control to the user without waiting for completion, we have to inject timings
+ double sleeptime = 0.0;
+ if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
+ //isend and send timings may be different
+ sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
+ }
+
+ if(sleeptime != 0.0){
+ simcall_process_sleep(sleeptime);
+ XBT_DEBUG("sending size of %zu : sleep %f ", request->size, smpi_os(request->size));
+ }
+
+ xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
+ xbt_mutex_acquire(mut);
+
if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
mailbox = smpi_process_remote_mailbox(receiver);
XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
if (request->flags & RMA || request->size < sg_cfg_get_int("smpi/async_small_thres")) { // eager mode
mailbox = smpi_process_remote_mailbox(receiver);
XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
- smx_action_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
+ smx_synchro_t action = simcall_comm_iprobe(mailbox, 1,request->dst, request->tag, &match_send, (void*)request);
- && ((char*)request->buf >= start_data_exe)
- && ((char*)request->buf < start_data_exe + size_data_exe )){
+ && ((char*) request->buf >= smpi_start_data_exe)
+ && ((char*)request->buf < smpi_start_data_exe + smpi_size_data_exe )){
buf = xbt_malloc(request->size);
memcpy(buf,oldbuf,request->size);
XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
buf = xbt_malloc(request->size);
memcpy(buf,oldbuf,request->size);
XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
- //if we are giving back the control to the user without waiting for completion, we have to inject timings
- double sleeptime = 0.0;
- if(request->detached || (request->flags & (ISEND|SSEND))){// issend should be treated as isend
- //isend and send timings may be different
- sleeptime = (request->flags & ISEND)? smpi_ois(request->size) : smpi_os(request->size);
- }
MPI_Datatype datatype = req->old_type;
if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
MPI_Datatype datatype = req->old_type;
if((req->flags & ACCUMULATE) || (datatype->has_subtype == 1)){
- && ((char*)req->old_buf >= start_data_exe)
- && ((char*)req->old_buf < start_data_exe + size_data_exe )
+ && ((char*)req->old_buf >= smpi_start_data_exe)
+ && ((char*)req->old_buf < smpi_start_data_exe + smpi_size_data_exe )
){
XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
smpi_switch_data_segment(smpi_process_index());
){
XBT_VERB("Privatization : We are unserializing to a zone in global memory - Switch data segment ");
smpi_switch_data_segment(smpi_process_index());
- //to avoid deadlocks
- //multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
+ // to avoid deadlocks if used as a break condition, such as
+ // while (MPI_Test(request, flag, status) && flag) {
+ // }
+ // because the time will not normally advance when only calls to MPI_Test
+ // are made -> deadlock
+ // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
map = xbt_new(int, count);
size = 0;
for(i = 0; i < count; i++) {
map = xbt_new(int, count);
size = 0;
for(i = 0; i < count; i++) {
MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
comm, PERSISTENT | RECV);
MPI_Request request =build_request(NULL, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : smpi_group_index(smpi_comm_group(comm), source), smpi_comm_rank(comm), tag,
comm, PERSISTENT | RECV);
- //to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
- //multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
+ // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
+ // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
+ // multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
static int nsleeps = 1;
if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
// behave like a receive, but don't do it
static int nsleeps = 1;
if(smpi_iprobe_sleep > 0) simcall_process_sleep(nsleeps*smpi_iprobe_sleep);
// behave like a receive, but don't do it
print_request("New iprobe", request);
// We have to test both mailboxes as we don't know if we will receive one one or another
print_request("New iprobe", request);
// We have to test both mailboxes as we don't know if we will receive one one or another
- if (sg_cfg_get_int("smpi/async_small_thres")>0){
- mailbox = smpi_process_mailbox_small();
- XBT_DEBUG("trying to probe the perm recv mailbox");
- request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
- }
- if (request->action==NULL){
- mailbox = smpi_process_mailbox();
- XBT_DEBUG("trying to probe the other mailbox");
- request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
- }
+ if (sg_cfg_get_int("smpi/async_small_thres")>0){
+ mailbox = smpi_process_mailbox_small();
+ XBT_DEBUG("trying to probe the perm recv mailbox");
+ request->action = simcall_comm_iprobe(mailbox, 0, request->src, request->tag, &match_recv, (void*)request);
+ }
+ if (request->action==NULL){
+ mailbox = smpi_process_mailbox();
+ XBT_DEBUG("trying to probe the other mailbox");
+ request->action = simcall_comm_iprobe(mailbox, 0, request->src,request->tag, &match_recv, (void*)request);
+ }
MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
*flag = 1;
if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
MPI_Request req = (MPI_Request)SIMIX_comm_get_src_data(request->action);
*flag = 1;
if(status != MPI_STATUS_IGNORE && !(req->flags & PREPARED)) {
status->MPI_SOURCE = smpi_group_rank(smpi_comm_group(comm), req->src);
- status->MPI_TAG = req->tag;
- status->MPI_ERROR = MPI_SUCCESS;
- status->count = req->real_size;
+ status->MPI_TAG = req->tag;
+ status->MPI_ERROR = MPI_SUCCESS;
+ status->count = req->real_size;
if ((*request)->action != NULL) { // this is not a detached send
simcall_comm_wait((*request)->action, -1.0);
if ((*request)->action != NULL) { // this is not a detached send
simcall_comm_wait((*request)->action, -1.0);
map = xbt_new(int, count);
size = 0;
XBT_DEBUG("Wait for one of %d", count);
map = xbt_new(int, count);
size = 0;
XBT_DEBUG("Wait for one of %d", count);
mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
recvcounts[rank], datatype, 0, comm);
xbt_free(displs);
mpi_coll_reduce_fun(sendbuf, tmpbuf, count, datatype, op, 0, comm);
smpi_mpi_scatterv(tmpbuf, recvcounts, displs, datatype, recvbuf,
recvcounts[rank], datatype, 0, comm);
xbt_free(displs);
}
void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
}
void smpi_mpi_gatherv(void *sendbuf, int sendcount, MPI_Datatype sendtype,
smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
}
smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
}
requests[index] =
smpi_irecv_init(tmpbufs[index], count, datatype, src,
system_tag, comm);
requests[index] =
smpi_irecv_init(tmpbufs[index], count, datatype, src,
system_tag, comm);
if(op) /* op can be MPI_OP_NULL that does nothing */
smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
}
if(op) /* op can be MPI_OP_NULL that does nothing */
smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
}
for(other = 0; other < rank; other++) {
// FIXME: possibly overkill we we have contiguous/noncontiguous data
// mapping...
for(other = 0; other < rank; other++) {
// FIXME: possibly overkill we we have contiguous/noncontiguous data
// mapping...
requests[index] =
smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
comm);
requests[index] =
smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
comm);
for(other = 0; other < rank; other++) {
// FIXME: possibly overkill we we have contiguous/noncontiguous data
// mapping...
for(other = 0; other < rank; other++) {
// FIXME: possibly overkill we we have contiguous/noncontiguous data
// mapping...
requests[index] =
smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
comm);
requests[index] =
smpi_irecv_init(tmpbufs[index], count, datatype, other, system_tag,
comm);