1 /* Copyright (c) 2007-2015. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include <xbt/config.hpp>
11 #include "xbt/replay.h"
13 #include "src/mc/mc_replay.h"
14 #include "src/simix/smx_private.h"
15 #include "simgrid/sg_config.h"
16 #include "smpi/smpi_utils.hpp"
17 #include <simgrid/s4u/host.hpp>
19 #include "src/kernel/activity/SynchroComm.hpp"
21 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
23 static simgrid::config::Flag<double> smpi_iprobe_sleep(
24 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
25 static simgrid::config::Flag<double> smpi_test_sleep(
26 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
28 std::vector<s_smpi_factor_t> smpi_os_values;
29 std::vector<s_smpi_factor_t> smpi_or_values;
30 std::vector<s_smpi_factor_t> smpi_ois_values;
32 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
34 static double smpi_os(size_t size)
36 if (smpi_os_values.empty()) {
37 smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
39 double current=smpi_os_values.empty()?0.0:smpi_os_values[0].values[0]+smpi_os_values[0].values[1]*size;
40 // Iterate over all the sections that were specified and find the right
41 // value. (fact.factor represents the interval sizes; we want to find the
42 // section that has fact.factor <= size and no other such fact.factor <= size)
43 // Note: parse_factor() (used before) already sorts the vector we iterate over!
44 for (auto& fact : smpi_os_values) {
45 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
46 XBT_DEBUG("os : %zu <= %zu return %.10f", size, fact.factor, current);
49 // If the next section is too large, the current section must be used.
50 // Hence, save the cost, as we might have to use it.
51 current = fact.values[0]+fact.values[1]*size;
54 XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current);
59 static double smpi_ois(size_t size)
61 if (smpi_ois_values.empty()) {
62 smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
64 double current=smpi_ois_values.empty()?0.0:smpi_ois_values[0].values[0]+smpi_ois_values[0].values[1]*size;
65 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
66 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
67 // Note: parse_factor() (used before) already sorts the vector we iterate over!
68 for (auto& fact : smpi_ois_values) {
69 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
70 XBT_DEBUG("ois : %zu <= %zu return %.10f", size, fact.factor, current);
73 // If the next section is too large, the current section must be used.
74 // Hence, save the cost, as we might have to use it.
75 current = fact.values[0]+fact.values[1]*size;
78 XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current);
83 static double smpi_or(size_t size)
85 if (smpi_or_values.empty()) {
86 smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
89 double current=smpi_or_values.empty()?0.0:smpi_or_values.front().values[0]+smpi_or_values.front().values[1]*size;
91 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
92 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
93 // Note: parse_factor() (used before) already sorts the vector we iterate over!
94 for (auto fact : smpi_or_values) {
95 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
96 XBT_DEBUG("or : %zu <= %zu return %.10f", size, fact.factor, current);
99 // If the next section is too large, the current section must be used.
100 // Hence, save the cost, as we might have to use it.
101 current=fact.values[0]+fact.values[1]*size;
104 XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current);
113 Request::Request(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm, unsigned flags) : buf_(buf), old_type_(datatype), src_(src), dst_(dst), tag_(tag), comm_(comm), flags_(flags)
115 void *old_buf = nullptr;
117 if((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED)){
118 // This part handles the problem of non-contiguous memory
120 buf_ = count==0 ? nullptr : xbt_malloc(count*datatype->size());
121 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
122 datatype->serialize(old_buf, buf_, count);
125 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
127 size_ = datatype->size() * count;
132 detached_sender_ = nullptr;
137 if (flags & PERSISTENT)
145 //Request::destroy(void* request)
147 // MPI_Request req = static_cast<MPI_Request>(request);
151 MPI_Comm Request::comm(){
167 int Request::flags(){
171 size_t Request::size(){
175 size_t Request::real_size(){
180 void Request::unuse(MPI_Request* request)
182 if((*request) != MPI_REQUEST_NULL){
183 (*request)->refcount_--;
184 if((*request)->refcount_<0) xbt_die("wrong refcount");
186 if((*request)->refcount_==0){
187 (*request)->old_type_->unuse();
188 (*request)->comm_->unuse();
189 (*request)->print_request("Destroying");
191 *request = MPI_REQUEST_NULL;
193 (*request)->print_request("Decrementing");
196 xbt_die("freeing an already free request");
201 int Request::match_recv(void* a, void* b, smx_activity_t ignored) {
202 MPI_Request ref = static_cast<MPI_Request>(a);
203 MPI_Request req = static_cast<MPI_Request>(b);
204 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
206 xbt_assert(ref, "Cannot match recv against null reference");
207 xbt_assert(req, "Cannot match recv against null request");
208 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
209 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
210 //we match, we can transfer some values
211 if(ref->src_ == MPI_ANY_SOURCE)
212 ref->real_src_ = req->src_;
213 if(ref->tag_ == MPI_ANY_TAG)
214 ref->real_tag_ = req->tag_;
215 if(ref->real_size_ < req->real_size_)
217 if(req->detached_==1)
218 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
219 XBT_DEBUG("match succeeded");
224 int Request::match_send(void* a, void* b,smx_activity_t ignored) {
225 MPI_Request ref = static_cast<MPI_Request>(a);
226 MPI_Request req = static_cast<MPI_Request>(b);
227 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
228 xbt_assert(ref, "Cannot match send against null reference");
229 xbt_assert(req, "Cannot match send against null request");
231 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
232 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
233 if(req->src_ == MPI_ANY_SOURCE)
234 req->real_src_ = ref->src_;
235 if(req->tag_ == MPI_ANY_TAG)
236 req->real_tag_ = ref->tag_;
237 if(req->real_size_ < ref->real_size_)
239 if(ref->detached_==1)
240 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
241 XBT_DEBUG("match succeeded");
247 void Request::print_request(const char *message)
249 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
250 message, this, buf_, size_, src_, dst_, tag_, flags_);
254 /* factories, to hide the internal flags from the caller */
255 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
257 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
258 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
259 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
263 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
265 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
266 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
267 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
271 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
273 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
274 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process_index(),
275 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
280 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
283 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
285 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
286 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
288 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
289 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
295 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
297 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
298 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
299 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
300 smpi_process_index(), tag, comm, PERSISTENT | RECV | PREPARED);
304 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
307 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
309 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
310 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
312 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
313 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
319 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
321 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
322 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
323 comm->group()->index(src), smpi_process_index(), tag,
324 comm, PERSISTENT | RECV | PREPARED);
328 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
330 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
331 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
332 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
337 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
339 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
340 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
341 comm->group()->index(dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
347 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
349 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
350 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
351 comm->group()->index(src), smpi_process_index(), tag, comm,
352 NON_PERSISTENT | RECV);
357 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
359 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
360 request = irecv(buf, count, datatype, src, tag, comm);
361 wait(&request,status);
365 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
367 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
368 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
369 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
372 wait(&request, MPI_STATUS_IGNORE);
376 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
378 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
379 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process_index(),
380 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
383 wait(&request,MPI_STATUS_IGNORE);
389 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
390 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
391 MPI_Comm comm, MPI_Status * status)
393 MPI_Request requests[2];
395 int myid=smpi_process_index();
396 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
397 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
400 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
401 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
402 startall(2, requests);
403 waitall(2, requests, stats);
406 if(status != MPI_STATUS_IGNORE) {
407 // Copy receive status
414 void Request::start()
416 smx_mailbox_t mailbox;
418 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
423 if ((flags_ & RECV) != 0) {
424 this->print_request("New recv");
426 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
428 xbt_mutex_t mut = smpi_process_mailboxes_mutex();
429 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
430 xbt_mutex_acquire(mut);
432 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
433 mailbox = smpi_process_mailbox();
435 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
436 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
437 //begin with the more appropriate one : the small one.
438 mailbox = smpi_process_mailbox_small();
439 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
440 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv,
441 static_cast<void*>(this));
443 if (action == nullptr) {
444 mailbox = smpi_process_mailbox();
445 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
446 action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
447 if (action == nullptr) {
448 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
449 mailbox = smpi_process_mailbox_small();
452 XBT_DEBUG("yes there was something for us in the large mailbox");
455 mailbox = smpi_process_mailbox_small();
456 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
457 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
459 if (action == nullptr) {
460 XBT_DEBUG("No, nothing in the permanent receive mailbox");
461 mailbox = smpi_process_mailbox();
463 XBT_DEBUG("yes there was something for us in the small mailbox");
467 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
469 action_ = simcall_comm_irecv(SIMIX_process_self(), mailbox, buf_, &real_size_, &match_recv,
470 ! smpi_process_get_replaying()? smpi_comm_copy_data_callback
471 : &smpi_comm_null_copy_buffer_callback, this, -1.0);
472 XBT_DEBUG("recv simcall posted");
474 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
475 xbt_mutex_release(mut);
476 } else { /* the RECV flag was not set, so this is a send */
480 if (TRACE_smpi_view_internals()) {
481 TRACE_smpi_send(rank, rank, receiver, tag_, size_);
483 this->print_request("New send");
486 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
487 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
488 void *oldbuf = nullptr;
490 XBT_DEBUG("Send request %p is detached", this);
492 if(!(old_type_->flags() & DT_FLAG_DERIVED)){
494 if (!smpi_process_get_replaying() && oldbuf != nullptr && size_!=0){
495 if((smpi_privatize_global_variables != 0)
496 && (static_cast<char*>(buf_) >= smpi_start_data_exe)
497 && (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
498 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
499 smpi_switch_data_segment(src_);
501 buf = xbt_malloc(size_);
502 memcpy(buf,oldbuf,size_);
503 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
508 //if we are giving back the control to the user without waiting for completion, we have to inject timings
509 double sleeptime = 0.0;
510 if(detached_ != 0 || ((flags_ & (ISEND|SSEND)) != 0)){// issend should be treated as isend
511 //isend and send timings may be different
512 sleeptime = ((flags_ & ISEND) != 0) ? smpi_ois(size_) : smpi_os(size_);
516 simcall_process_sleep(sleeptime);
517 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
520 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
522 xbt_mutex_t mut=smpi_process_remote_mailboxes_mutex(receiver);
524 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
525 xbt_mutex_acquire(mut);
527 if (!(async_small_thresh != 0 || (flags_ & RMA) !=0)) {
528 mailbox = smpi_process_remote_mailbox(receiver);
529 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
530 mailbox = smpi_process_remote_mailbox(receiver);
531 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
532 smx_activity_t action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send,
533 static_cast<void*>(this));
534 if (action == nullptr) {
535 if ((flags_ & SSEND) == 0){
536 mailbox = smpi_process_remote_mailbox_small(receiver);
537 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
539 mailbox = smpi_process_remote_mailbox_small(receiver);
540 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
541 action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send, static_cast<void*>(this));
542 if (action == nullptr) {
543 XBT_DEBUG("No, we are first, send to large mailbox");
544 mailbox = smpi_process_remote_mailbox(receiver);
548 XBT_DEBUG("Yes there was something for us in the large mailbox");
551 mailbox = smpi_process_remote_mailbox(receiver);
552 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
555 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
557 action_ = simcall_comm_isend(SIMIX_process_from_PID(src_+1), mailbox, size_, -1.0,
558 buf, real_size_, &match_send,
559 &xbt_free_f, // how to free the userdata if a detached send fails
560 !smpi_process_get_replaying() ? smpi_comm_copy_data_callback
561 : &smpi_comm_null_copy_buffer_callback, this,
562 // detach if msg size < eager/rdv switch limit
564 XBT_DEBUG("send simcall posted");
566 /* FIXME: detached sends are not traceable (action_ == nullptr) */
567 if (action_ != nullptr)
568 simcall_set_category(action_, TRACE_internal_smpi_get_category());
570 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
571 xbt_mutex_release(mut);
576 void Request::startall(int count, MPI_Request * requests)
578 if(requests== nullptr)
581 for(int i = 0; i < count; i++) {
582 requests[i]->start();
586 int Request::test(MPI_Request * request, MPI_Status * status) {
587 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
589 // to avoid deadlocks if used as a break condition, such as
590 // while (MPI_Test(request, flag, status) && flag) {
592 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
593 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
594 static int nsleeps = 1;
595 if(smpi_test_sleep > 0)
596 simcall_process_sleep(nsleeps*smpi_test_sleep);
598 smpi_empty_status(status);
600 if (((*request)->flags_ & PREPARED) == 0) {
601 if ((*request)->action_ != nullptr)
602 flag = simcall_comm_test((*request)->action_);
604 finish_wait(request,status);
605 nsleeps=1;//reset the number of sleeps we will do next time
606 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT)==0)
607 *request = MPI_REQUEST_NULL;
608 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
616 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
622 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
624 for(i = 0; i < incount; i++) {
625 if((requests[i] != MPI_REQUEST_NULL)) {
626 if(test(&requests[i], pstat)) {
629 if(status != MPI_STATUSES_IGNORE) {
632 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags_ & NON_PERSISTENT)
633 requests[i]=MPI_REQUEST_NULL;
639 if(count_dead==incount)
640 return MPI_UNDEFINED;
645 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
647 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
648 comms.reserve(count);
653 *index = MPI_UNDEFINED;
655 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
656 for(i = 0; i < count; i++) {
657 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && !(requests[i]->flags_ & PREPARED)) {
658 comms.push_back(requests[i]->action_);
663 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
664 static int nsleeps = 1;
665 if(smpi_test_sleep > 0)
666 simcall_process_sleep(nsleeps*smpi_test_sleep);
668 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
669 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
671 finish_wait(&requests[*index],status);
674 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
675 requests[*index] = MPI_REQUEST_NULL;
681 //all requests are null or inactive, return true
683 smpi_empty_status(status);
690 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
693 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
695 for(int i=0; i<count; i++){
696 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED)) {
697 if (test(&requests[i], pstat)!=1){
700 requests[i]=MPI_REQUEST_NULL;
703 smpi_empty_status(pstat);
705 if(status != MPI_STATUSES_IGNORE) {
715 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
717 //FIXME find another way to avoid busy waiting ?
718 // the issue here is that we have to wait on a nonexistent comm
720 iprobe(source, tag, comm, &flag, status);
721 XBT_DEBUG("Busy Waiting on probing : %d", flag);
725 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
726 MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
727 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
729 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
730 // (especially when used as a break condition, such as while(MPI_Iprobe(...)) ... )
731 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
732 // (This can speed up the execution of certain applications by an order of magnitude, such as HPL)
733 static int nsleeps = 1;
734 double speed = simgrid::s4u::Actor::self()->host()->speed();
735 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
736 if (smpi_iprobe_sleep > 0) {
737 smx_activity_t iprobe_sleep = simcall_execution_start("iprobe", /* flops to executek*/nsleeps*smpi_iprobe_sleep*speed*maxrate, /* priority */1.0, /* performance bound */maxrate*speed);
738 simcall_execution_wait(iprobe_sleep);
740 // behave like a receive, but don't do it
741 smx_mailbox_t mailbox;
743 request->print_request("New iprobe");
744 // We have to test both mailboxes as we don't know if we will receive one one or another
745 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
746 mailbox = smpi_process_mailbox_small();
747 XBT_DEBUG("Trying to probe the perm recv mailbox");
748 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_, request->tag_, &match_recv,
749 static_cast<void*>(request));
752 if (request->action_ == nullptr){
753 mailbox = smpi_process_mailbox();
754 XBT_DEBUG("trying to probe the other mailbox");
755 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_,request->tag_, &match_recv,
756 static_cast<void*>(request));
759 if (request->action_ != nullptr){
760 simgrid::kernel::activity::Comm *sync_comm = static_cast<simgrid::kernel::activity::Comm*>(request->action_);
761 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
763 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
764 status->MPI_SOURCE = comm->group()->rank(req->src_);
765 status->MPI_TAG = req->tag_;
766 status->MPI_ERROR = MPI_SUCCESS;
767 status->count = req->real_size_;
769 nsleeps = 1;//reset the number of sleeps we will do next time
773 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
780 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
782 MPI_Request req = *request;
783 smpi_empty_status(status);
785 if(!((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)){
786 if(status != MPI_STATUS_IGNORE) {
787 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
788 status->MPI_SOURCE = req->comm_->group()->rank(src);
789 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
790 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
791 // this handles the case were size in receive differs from size in send
792 status->count = req->real_size_;
795 req->print_request("Finishing");
796 MPI_Datatype datatype = req->old_type_;
798 if(((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED)){
799 if (!smpi_process_get_replaying()){
800 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
801 && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
802 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
803 smpi_switch_data_segment(smpi_process_index());
807 if(datatype->flags() & DT_FLAG_DERIVED){
808 // This part handles the problem of non-contignous memory the unserialization at the reception
809 if((req->flags_ & RECV) && datatype->size()!=0)
810 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
812 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
813 int n =req->real_size_/datatype->size();
814 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
820 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
821 int rank = smpi_process_index();
822 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
823 TRACE_smpi_recv(rank, src_traced, rank,req->tag_);
825 if(req->detached_sender_ != nullptr){
826 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
827 double sleeptime = smpi_or(req->real_size_);
829 simcall_process_sleep(sleeptime);
830 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
832 unuse(&(req->detached_sender_));
834 if(req->flags_ & PERSISTENT)
835 req->action_ = nullptr;
836 req->flags_ |= FINISHED;
841 void Request::wait(MPI_Request * request, MPI_Status * status)
843 (*request)->print_request("Waiting");
844 if ((*request)->flags_ & PREPARED) {
845 smpi_empty_status(status);
849 if ((*request)->action_ != nullptr)
850 // this is not a detached send
851 simcall_comm_wait((*request)->action_, -1.0);
853 finish_wait(request,status);
854 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
855 *request = MPI_REQUEST_NULL;
858 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
860 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
863 int index = MPI_UNDEFINED;
867 // Wait for a request to complete
868 xbt_dynar_init(&comms, sizeof(smx_activity_t), nullptr);
869 map = xbt_new(int, count);
870 XBT_DEBUG("Wait for one of %d", count);
871 for(i = 0; i < count; i++) {
872 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED) && !(requests[i]->flags_ & FINISHED)) {
873 if (requests[i]->action_ != nullptr) {
874 XBT_DEBUG("Waiting any %p ", requests[i]);
875 xbt_dynar_push(&comms, &requests[i]->action_);
879 // This is a finished detached request, let's return this one
880 size = 0; // so we free the dynar but don't do the waitany call
882 finish_wait(&requests[i], status); // cleanup if refcount = 0
883 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
884 requests[i] = MPI_REQUEST_NULL; // set to null
890 i = simcall_comm_waitany(&comms, -1);
892 // not MPI_UNDEFINED, as this is a simix return code
895 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
896 if ((requests[index] == MPI_REQUEST_NULL)
897 || (!((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))){
898 finish_wait(&requests[index],status);
899 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
900 requests[index] = MPI_REQUEST_NULL;
907 xbt_dynar_free_data(&comms);
911 if (index==MPI_UNDEFINED)
912 smpi_empty_status(status);
917 static int sort_accumulates(MPI_Request a, MPI_Request b)
919 return (a->tag() < b->tag());
922 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
924 std::vector<MPI_Request> accumulates;
927 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
928 int retvalue = MPI_SUCCESS;
929 //tag invalid requests in the set
930 if (status != MPI_STATUSES_IGNORE) {
931 for (int c = 0; c < count; c++) {
932 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
933 smpi_empty_status(&status[c]);
934 } else if (requests[c]->src_ == MPI_PROC_NULL) {
935 smpi_empty_status(&status[c]);
936 status[c].MPI_SOURCE = MPI_PROC_NULL;
940 for (int c = 0; c < count; c++) {
941 if (MC_is_active() || MC_record_replay_is_active()) {
942 wait(&requests[c],pstat);
945 index = waitany(count, requests, pstat);
946 if (index == MPI_UNDEFINED)
949 if (requests[index] != MPI_REQUEST_NULL
950 && (requests[index]->flags_ & RECV)
951 && (requests[index]->flags_ & ACCUMULATE))
952 accumulates.push_back(requests[index]);
953 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
954 requests[index] = MPI_REQUEST_NULL;
956 if (status != MPI_STATUSES_IGNORE) {
957 status[index] = *pstat;
958 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
959 retvalue = MPI_ERR_IN_STATUS;
963 if (!accumulates.empty()) {
964 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
965 for (auto req : accumulates) {
966 finish_wait(&req, status);
973 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
979 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
981 for(i = 0; i < incount; i++)
983 index=waitany(incount, requests, pstat);
984 if(index!=MPI_UNDEFINED){
985 indices[count] = index;
987 if(status != MPI_STATUSES_IGNORE) {
988 status[index] = *pstat;
990 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
991 requests[index]=MPI_REQUEST_NULL;
993 return MPI_UNDEFINED;