1 /* Copyright (c) 2007-2017. The SimGrid Team. All rights reserved. */
3 /* This program is free software; you can redistribute it and/or modify it
4 * under the terms of the license (GNU LGPL) which comes with this package. */
6 #include <xbt/config.hpp>
10 #include "src/mc/mc_replay.h"
11 #include "src/simix/smx_private.h"
12 #include "simgrid/sg_config.h"
13 #include "smpi/smpi_utils.hpp"
14 #include <simgrid/s4u/host.hpp>
15 #include "src/kernel/activity/SynchroComm.hpp"
17 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
19 static simgrid::config::Flag<double> smpi_iprobe_sleep(
20 "smpi/iprobe", "Minimum time to inject inside a call to MPI_Iprobe", 1e-4);
21 static simgrid::config::Flag<double> smpi_test_sleep(
22 "smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
24 std::vector<s_smpi_factor_t> smpi_os_values;
25 std::vector<s_smpi_factor_t> smpi_or_values;
26 std::vector<s_smpi_factor_t> smpi_ois_values;
28 extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
30 static double smpi_os(size_t size)
32 if (smpi_os_values.empty()) {
33 smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
35 double current=smpi_os_values.empty()?0.0:smpi_os_values[0].values[0]+smpi_os_values[0].values[1]*size;
36 // Iterate over all the sections that were specified and find the right
37 // value. (fact.factor represents the interval sizes; we want to find the
38 // section that has fact.factor <= size and no other such fact.factor <= size)
39 // Note: parse_factor() (used before) already sorts the vector we iterate over!
40 for (auto& fact : smpi_os_values) {
41 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
42 XBT_DEBUG("os : %zu <= %zu return %.10f", size, fact.factor, current);
45 // If the next section is too large, the current section must be used.
46 // Hence, save the cost, as we might have to use it.
47 current = fact.values[0]+fact.values[1]*size;
50 XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current);
55 static double smpi_ois(size_t size)
57 if (smpi_ois_values.empty()) {
58 smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
60 double current=smpi_ois_values.empty()?0.0:smpi_ois_values[0].values[0]+smpi_ois_values[0].values[1]*size;
61 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
62 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
63 // Note: parse_factor() (used before) already sorts the vector we iterate over!
64 for (auto& fact : smpi_ois_values) {
65 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
66 XBT_DEBUG("ois : %zu <= %zu return %.10f", size, fact.factor, current);
69 // If the next section is too large, the current section must be used.
70 // Hence, save the cost, as we might have to use it.
71 current = fact.values[0]+fact.values[1]*size;
74 XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current);
79 static double smpi_or(size_t size)
81 if (smpi_or_values.empty()) {
82 smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
85 double current=smpi_or_values.empty()?0.0:smpi_or_values.front().values[0]+smpi_or_values.front().values[1]*size;
87 // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
88 // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
89 // Note: parse_factor() (used before) already sorts the vector we iterate over!
90 for (auto fact : smpi_or_values) {
91 if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
92 XBT_DEBUG("or : %zu <= %zu return %.10f", size, fact.factor, current);
95 // If the next section is too large, the current section must be used.
96 // Hence, save the cost, as we might have to use it.
97 current=fact.values[0]+fact.values[1]*size;
100 XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current);
108 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)
110 void *old_buf = nullptr;
111 if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(buf_))){
112 // This part handles the problem of non-contiguous memory
114 buf_ = count==0 ? nullptr : xbt_malloc(count*datatype->size());
115 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
116 datatype->serialize(old_buf, buf_, count);
119 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
121 size_ = datatype->size() * count;
126 detached_sender_ = nullptr;
131 if (flags & PERSISTENT)
138 MPI_Comm Request::comm(){
154 int Request::flags(){
158 int Request::detached(){
162 size_t Request::size(){
166 size_t Request::real_size(){
170 void Request::unref(MPI_Request* request)
172 if((*request) != MPI_REQUEST_NULL){
173 (*request)->refcount_--;
174 if((*request)->refcount_<0) xbt_die("wrong refcount");
175 if((*request)->refcount_==0){
176 Datatype::unref((*request)->old_type_);
177 Comm::unref((*request)->comm_);
178 (*request)->print_request("Destroying");
180 *request = MPI_REQUEST_NULL;
182 (*request)->print_request("Decrementing");
185 xbt_die("freeing an already free request");
189 int Request::match_recv(void* a, void* b, smx_activity_t ignored) {
190 MPI_Request ref = static_cast<MPI_Request>(a);
191 MPI_Request req = static_cast<MPI_Request>(b);
192 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
194 xbt_assert(ref, "Cannot match recv against null reference");
195 xbt_assert(req, "Cannot match recv against null request");
196 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
197 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
198 //we match, we can transfer some values
199 if(ref->src_ == MPI_ANY_SOURCE)
200 ref->real_src_ = req->src_;
201 if(ref->tag_ == MPI_ANY_TAG)
202 ref->real_tag_ = req->tag_;
203 if(ref->real_size_ < req->real_size_)
205 if(req->detached_==1)
206 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
207 XBT_DEBUG("match succeeded");
212 int Request::match_send(void* a, void* b,smx_activity_t ignored) {
213 MPI_Request ref = static_cast<MPI_Request>(a);
214 MPI_Request req = static_cast<MPI_Request>(b);
215 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
216 xbt_assert(ref, "Cannot match send against null reference");
217 xbt_assert(req, "Cannot match send against null request");
219 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
220 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
221 if(req->src_ == MPI_ANY_SOURCE)
222 req->real_src_ = ref->src_;
223 if(req->tag_ == MPI_ANY_TAG)
224 req->real_tag_ = ref->tag_;
225 if(req->real_size_ < ref->real_size_)
227 if(ref->detached_==1)
228 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
229 XBT_DEBUG("match succeeded");
235 void Request::print_request(const char *message)
237 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
238 message, this, buf_, size_, src_, dst_, tag_, flags_);
242 /* factories, to hide the internal flags from the caller */
243 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
246 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
247 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
250 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
252 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
253 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
256 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
258 return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
259 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
263 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
266 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
268 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
269 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
271 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
272 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
278 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
280 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
281 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
282 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
285 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
288 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
290 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
291 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
293 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
294 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
300 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
302 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
303 comm->group()->index(src), smpi_process()->index(), tag,
304 comm, PERSISTENT | RECV | PREPARED);
307 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
309 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
310 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
311 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
316 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
318 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
319 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
320 comm->group()->index(dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
326 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
328 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
329 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
330 comm->group()->index(src), smpi_process()->index(), tag, comm,
331 NON_PERSISTENT | RECV);
336 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
338 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
339 request = irecv(buf, count, datatype, src, tag, comm);
340 wait(&request,status);
344 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
346 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
347 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
348 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
351 wait(&request, MPI_STATUS_IGNORE);
355 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
357 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
358 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
359 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
362 wait(&request,MPI_STATUS_IGNORE);
366 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
367 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
368 MPI_Comm comm, MPI_Status * status)
370 MPI_Request requests[2];
372 int myid=smpi_process()->index();
373 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
374 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
377 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
378 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
379 startall(2, requests);
380 waitall(2, requests, stats);
383 if(status != MPI_STATUS_IGNORE) {
384 // Copy receive status
389 void Request::start()
391 smx_mailbox_t mailbox;
393 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
398 if ((flags_ & RECV) != 0) {
399 this->print_request("New recv");
401 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
403 xbt_mutex_t mut = smpi_process()->mailboxes_mutex();
404 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
405 xbt_mutex_acquire(mut);
407 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
408 mailbox = smpi_process()->mailbox();
410 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
411 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
412 //begin with the more appropriate one : the small one.
413 mailbox = smpi_process()->mailbox_small();
414 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
415 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv,
416 static_cast<void*>(this));
418 if (action == nullptr) {
419 mailbox = smpi_process()->mailbox();
420 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
421 action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
422 if (action == nullptr) {
423 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
424 mailbox = smpi_process()->mailbox_small();
427 XBT_DEBUG("yes there was something for us in the large mailbox");
430 mailbox = smpi_process()->mailbox_small();
431 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
432 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
434 if (action == nullptr) {
435 XBT_DEBUG("No, nothing in the permanent receive mailbox");
436 mailbox = smpi_process()->mailbox();
438 XBT_DEBUG("yes there was something for us in the small mailbox");
442 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
444 action_ = simcall_comm_irecv(SIMIX_process_self(), mailbox, buf_, &real_size_, &match_recv,
445 ! smpi_process()->replaying()? smpi_comm_copy_data_callback
446 : &smpi_comm_null_copy_buffer_callback, this, -1.0);
447 XBT_DEBUG("recv simcall posted");
449 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
450 xbt_mutex_release(mut);
451 } else { /* the RECV flag was not set, so this is a send */
455 if (TRACE_smpi_view_internals()) {
456 TRACE_smpi_send(rank, rank, receiver, tag_, size_);
458 this->print_request("New send");
461 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
462 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
463 void *oldbuf = nullptr;
465 XBT_DEBUG("Send request %p is detached", this);
467 if(!(old_type_->flags() & DT_FLAG_DERIVED)){
469 if (!smpi_process()->replaying() && oldbuf != nullptr && size_!=0){
470 if((smpi_privatize_global_variables != 0)
471 && (static_cast<char*>(buf_) >= smpi_start_data_exe)
472 && (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
473 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
474 smpi_switch_data_segment(src_);
476 buf = xbt_malloc(size_);
477 memcpy(buf,oldbuf,size_);
478 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
483 //if we are giving back the control to the user without waiting for completion, we have to inject timings
484 double sleeptime = 0.0;
485 if(detached_ != 0 || ((flags_ & (ISEND|SSEND)) != 0)){// issend should be treated as isend
486 //isend and send timings may be different
487 sleeptime = ((flags_ & ISEND) != 0) ? smpi_ois(size_) : smpi_os(size_);
491 simcall_process_sleep(sleeptime);
492 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
495 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
497 xbt_mutex_t mut=smpi_process_remote(receiver)->mailboxes_mutex();
499 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
500 xbt_mutex_acquire(mut);
502 if (!(async_small_thresh != 0 || (flags_ & RMA) !=0)) {
503 mailbox = smpi_process_remote(receiver)->mailbox();
504 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
505 mailbox = smpi_process_remote(receiver)->mailbox();
506 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
507 smx_activity_t action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send,
508 static_cast<void*>(this));
509 if (action == nullptr) {
510 if ((flags_ & SSEND) == 0){
511 mailbox = smpi_process_remote(receiver)->mailbox_small();
512 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
514 mailbox = smpi_process_remote(receiver)->mailbox_small();
515 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
516 action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send, static_cast<void*>(this));
517 if (action == nullptr) {
518 XBT_DEBUG("No, we are first, send to large mailbox");
519 mailbox = smpi_process_remote(receiver)->mailbox();
523 XBT_DEBUG("Yes there was something for us in the large mailbox");
526 mailbox = smpi_process_remote(receiver)->mailbox();
527 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
530 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
532 action_ = simcall_comm_isend(SIMIX_process_from_PID(src_+1), mailbox, size_, -1.0,
533 buf, real_size_, &match_send,
534 &xbt_free_f, // how to free the userdata if a detached send fails
535 !smpi_process()->replaying() ? smpi_comm_copy_data_callback
536 : &smpi_comm_null_copy_buffer_callback, this,
537 // detach if msg size < eager/rdv switch limit
539 XBT_DEBUG("send simcall posted");
541 /* FIXME: detached sends are not traceable (action_ == nullptr) */
542 if (action_ != nullptr)
543 simcall_set_category(action_, TRACE_internal_smpi_get_category());
544 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
545 xbt_mutex_release(mut);
549 void Request::startall(int count, MPI_Request * requests)
551 if(requests== nullptr)
554 for(int i = 0; i < count; i++) {
555 requests[i]->start();
559 int Request::test(MPI_Request * request, MPI_Status * status) {
560 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
561 // to avoid deadlocks if used as a break condition, such as
562 // while (MPI_Test(request, flag, status) && flag) dostuff...
563 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
564 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
565 static int nsleeps = 1;
566 if(smpi_test_sleep > 0)
567 simcall_process_sleep(nsleeps*smpi_test_sleep);
569 Status::empty(status);
571 if (((*request)->flags_ & PREPARED) == 0) {
572 if ((*request)->action_ != nullptr)
573 flag = simcall_comm_test((*request)->action_);
575 finish_wait(request,status);
576 nsleeps=1;//reset the number of sleeps we will do next time
577 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT)==0)
578 *request = MPI_REQUEST_NULL;
579 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
586 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
592 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
594 for(i = 0; i < incount; i++) {
595 if((requests[i] != MPI_REQUEST_NULL)) {
596 if(test(&requests[i], pstat)) {
599 if(status != MPI_STATUSES_IGNORE) {
602 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags_ & NON_PERSISTENT)
603 requests[i]=MPI_REQUEST_NULL;
609 if(count_dead==incount)
610 return MPI_UNDEFINED;
614 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
616 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
617 comms.reserve(count);
622 *index = MPI_UNDEFINED;
624 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
625 for(i = 0; i < count; i++) {
626 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && !(requests[i]->flags_ & PREPARED)) {
627 comms.push_back(requests[i]->action_);
632 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
633 static int nsleeps = 1;
634 if(smpi_test_sleep > 0)
635 simcall_process_sleep(nsleeps*smpi_test_sleep);
637 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
638 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
640 finish_wait(&requests[*index],status);
643 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
644 requests[*index] = MPI_REQUEST_NULL;
650 //all requests are null or inactive, return true
652 Status::empty(status);
658 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
661 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
663 for(int i=0; i<count; i++){
664 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED)) {
665 if (test(&requests[i], pstat)!=1){
668 requests[i]=MPI_REQUEST_NULL;
671 Status::empty(pstat);
673 if(status != MPI_STATUSES_IGNORE) {
680 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
682 //FIXME find another way to avoid busy waiting ?
683 // the issue here is that we have to wait on a nonexistent comm
685 iprobe(source, tag, comm, &flag, status);
686 XBT_DEBUG("Busy Waiting on probing : %d", flag);
690 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
691 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
692 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
693 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
694 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
695 static int nsleeps = 1;
696 double speed = simgrid::s4u::Actor::self()->host()->speed();
697 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
698 MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
699 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
700 if (smpi_iprobe_sleep > 0) {
701 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);
702 simcall_execution_wait(iprobe_sleep);
704 // behave like a receive, but don't do it
705 smx_mailbox_t mailbox;
707 request->print_request("New iprobe");
708 // We have to test both mailboxes as we don't know if we will receive one one or another
709 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
710 mailbox = smpi_process()->mailbox_small();
711 XBT_DEBUG("Trying to probe the perm recv mailbox");
712 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_, request->tag_, &match_recv,
713 static_cast<void*>(request));
716 if (request->action_ == nullptr){
717 mailbox = smpi_process()->mailbox();
718 XBT_DEBUG("trying to probe the other mailbox");
719 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_,request->tag_, &match_recv,
720 static_cast<void*>(request));
723 if (request->action_ != nullptr){
724 simgrid::kernel::activity::Comm *sync_comm = static_cast<simgrid::kernel::activity::Comm*>(request->action_);
725 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
727 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
728 status->MPI_SOURCE = comm->group()->rank(req->src_);
729 status->MPI_TAG = req->tag_;
730 status->MPI_ERROR = MPI_SUCCESS;
731 status->count = req->real_size_;
733 nsleeps = 1;//reset the number of sleeps we will do next time
737 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
743 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
745 MPI_Request req = *request;
746 Status::empty(status);
748 if(!((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)){
749 if(status != MPI_STATUS_IGNORE) {
750 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
751 status->MPI_SOURCE = req->comm_->group()->rank(src);
752 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
753 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
754 // this handles the case were size in receive differs from size in send
755 status->count = req->real_size_;
758 req->print_request("Finishing");
759 MPI_Datatype datatype = req->old_type_;
761 if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(req->old_buf_))){
763 if (!smpi_process()->replaying()){
764 if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
765 && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
766 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
767 smpi_switch_data_segment(smpi_process()->index());
771 if(datatype->flags() & DT_FLAG_DERIVED){
772 // This part handles the problem of non-contignous memory the unserialization at the reception
773 if((req->flags_ & RECV) && datatype->size()!=0)
774 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
776 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
777 int n =req->real_size_/datatype->size();
778 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
784 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
785 int rank = smpi_process()->index();
786 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
787 TRACE_smpi_recv(rank, src_traced, rank,req->tag_);
789 if(req->detached_sender_ != nullptr){
790 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
791 double sleeptime = smpi_or(req->real_size_);
793 simcall_process_sleep(sleeptime);
794 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
796 unref(&(req->detached_sender_));
798 if(req->flags_ & PERSISTENT)
799 req->action_ = nullptr;
800 req->flags_ |= FINISHED;
804 void Request::wait(MPI_Request * request, MPI_Status * status)
806 (*request)->print_request("Waiting");
807 if ((*request)->flags_ & PREPARED) {
808 Status::empty(status);
812 if ((*request)->action_ != nullptr)
813 // this is not a detached send
814 simcall_comm_wait((*request)->action_, -1.0);
816 finish_wait(request,status);
817 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
818 *request = MPI_REQUEST_NULL;
821 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
823 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
826 int index = MPI_UNDEFINED;
830 // Wait for a request to complete
831 xbt_dynar_init(&comms, sizeof(smx_activity_t), nullptr);
832 map = xbt_new(int, count);
833 XBT_DEBUG("Wait for one of %d", count);
834 for(i = 0; i < count; i++) {
835 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED) && !(requests[i]->flags_ & FINISHED)) {
836 if (requests[i]->action_ != nullptr) {
837 XBT_DEBUG("Waiting any %p ", requests[i]);
838 xbt_dynar_push(&comms, &requests[i]->action_);
842 // This is a finished detached request, let's return this one
843 size = 0; // so we free the dynar but don't do the waitany call
845 finish_wait(&requests[i], status); // cleanup if refcount = 0
846 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
847 requests[i] = MPI_REQUEST_NULL; // set to null
853 i = simcall_comm_waitany(&comms, -1);
855 // not MPI_UNDEFINED, as this is a simix return code
858 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
859 if ((requests[index] == MPI_REQUEST_NULL)
860 || (!((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))){
861 finish_wait(&requests[index],status);
862 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
863 requests[index] = MPI_REQUEST_NULL;
868 xbt_dynar_free_data(&comms);
872 if (index==MPI_UNDEFINED)
873 Status::empty(status);
878 static int sort_accumulates(MPI_Request a, MPI_Request b)
880 return (a->tag() > b->tag());
883 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
885 std::vector<MPI_Request> accumulates;
888 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
889 int retvalue = MPI_SUCCESS;
890 //tag invalid requests in the set
891 if (status != MPI_STATUSES_IGNORE) {
892 for (int c = 0; c < count; c++) {
893 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
894 Status::empty(&status[c]);
895 } else if (requests[c]->src_ == MPI_PROC_NULL) {
896 Status::empty(&status[c]);
897 status[c].MPI_SOURCE = MPI_PROC_NULL;
901 for (int c = 0; c < count; c++) {
902 if (MC_is_active() || MC_record_replay_is_active()) {
903 wait(&requests[c],pstat);
906 index = waitany(count, requests, pstat);
907 if (index == MPI_UNDEFINED)
910 if (requests[index] != MPI_REQUEST_NULL
911 && (requests[index]->flags_ & RECV)
912 && (requests[index]->flags_ & ACCUMULATE))
913 accumulates.push_back(requests[index]);
914 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
915 requests[index] = MPI_REQUEST_NULL;
917 if (status != MPI_STATUSES_IGNORE) {
918 status[index] = *pstat;
919 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
920 retvalue = MPI_ERR_IN_STATUS;
924 if (!accumulates.empty()) {
925 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
926 for (auto req : accumulates) {
927 finish_wait(&req, status);
934 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
940 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
942 for(i = 0; i < incount; i++)
944 index=waitany(incount, requests, pstat);
945 if(index!=MPI_UNDEFINED){
946 indices[count] = index;
948 if(status != MPI_STATUSES_IGNORE) {
949 status[index] = *pstat;
951 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
952 requests[index]=MPI_REQUEST_NULL;
954 return MPI_UNDEFINED;
960 MPI_Request Request::f2c(int id) {
962 if(id==MPI_FORTRAN_REQUEST_NULL)
963 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
964 return static_cast<MPI_Request>(xbt_dict_get(F2C::f2c_lookup(), get_key_id(key, id)));
967 int Request::add_f() {
968 if(F2C::f2c_lookup()==nullptr){
969 F2C::set_f2c_lookup(xbt_dict_new_homogeneous(nullptr));
972 xbt_dict_set(F2C::f2c_lookup(), get_key_id(key, F2C::f2c_id()), this, nullptr);
973 F2C::f2c_id_increment();
974 return F2C::f2c_id()-1;
977 void Request::free_f(int id) {
979 if(id!=MPI_FORTRAN_REQUEST_NULL)
980 xbt_dict_remove(F2C::f2c_lookup(), get_key_id(key, id));