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 // FIXME Handle the case of a partial shared malloc.
112 if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED))) { // && (!smpi_is_shared(buf_))){
113 // This part handles the problem of non-contiguous memory
118 buf_ = xbt_malloc(count*datatype->size());
119 if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
120 datatype->serialize(old_buf, buf_, count);
124 // This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
126 size_ = datatype->size() * count;
131 detached_sender_ = nullptr;
136 if (flags & PERSISTENT)
143 MPI_Comm Request::comm(){
159 int Request::flags(){
163 int Request::detached(){
167 size_t Request::size(){
171 size_t Request::real_size(){
175 void Request::unref(MPI_Request* request)
177 if((*request) != MPI_REQUEST_NULL){
178 (*request)->refcount_--;
179 if((*request)->refcount_<0) xbt_die("wrong refcount");
180 if((*request)->refcount_==0){
181 Datatype::unref((*request)->old_type_);
182 Comm::unref((*request)->comm_);
183 (*request)->print_request("Destroying");
185 *request = MPI_REQUEST_NULL;
187 (*request)->print_request("Decrementing");
190 xbt_die("freeing an already free request");
194 int Request::match_recv(void* a, void* b, smx_activity_t ignored) {
195 MPI_Request ref = static_cast<MPI_Request>(a);
196 MPI_Request req = static_cast<MPI_Request>(b);
197 XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
199 xbt_assert(ref, "Cannot match recv against null reference");
200 xbt_assert(req, "Cannot match recv against null request");
201 if((ref->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
202 && ((ref->tag_ == MPI_ANY_TAG && req->tag_ >=0) || req->tag_ == ref->tag_)){
203 //we match, we can transfer some values
204 if(ref->src_ == MPI_ANY_SOURCE)
205 ref->real_src_ = req->src_;
206 if(ref->tag_ == MPI_ANY_TAG)
207 ref->real_tag_ = req->tag_;
208 if(ref->real_size_ < req->real_size_)
210 if(req->detached_==1)
211 ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
212 XBT_DEBUG("match succeeded");
217 int Request::match_send(void* a, void* b,smx_activity_t ignored) {
218 MPI_Request ref = static_cast<MPI_Request>(a);
219 MPI_Request req = static_cast<MPI_Request>(b);
220 XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
221 xbt_assert(ref, "Cannot match send against null reference");
222 xbt_assert(req, "Cannot match send against null request");
224 if((req->src_ == MPI_ANY_SOURCE || req->src_ == ref->src_)
225 && ((req->tag_ == MPI_ANY_TAG && ref->tag_ >=0)|| req->tag_ == ref->tag_)){
226 if(req->src_ == MPI_ANY_SOURCE)
227 req->real_src_ = ref->src_;
228 if(req->tag_ == MPI_ANY_TAG)
229 req->real_tag_ = ref->tag_;
230 if(req->real_size_ < ref->real_size_)
232 if(ref->detached_==1)
233 req->detached_sender_=ref; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
234 XBT_DEBUG("match succeeded");
240 void Request::print_request(const char *message)
242 XBT_VERB("%s request %p [buf = %p, size = %zu, src = %d, dst = %d, tag = %d, flags = %x]",
243 message, this, buf_, size_, src_, dst_, tag_, flags_);
247 /* factories, to hide the internal flags from the caller */
248 MPI_Request Request::send_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
251 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
252 comm->group()->index(dst), tag, comm, PERSISTENT | SEND | PREPARED);
255 MPI_Request Request::ssend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
257 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
258 comm->group()->index(dst), tag, comm, PERSISTENT | SSEND | SEND | PREPARED);
261 MPI_Request Request::isend_init(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
263 return new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, smpi_process()->index(),
264 comm->group()->index(dst), tag,comm, PERSISTENT | ISEND | SEND | PREPARED);
268 MPI_Request Request::rma_send_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
271 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
273 request = new Request(buf==MPI_BOTTOM ? nullptr : buf , count, datatype, src, dst, tag,
274 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED);
276 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
277 comm, RMA | NON_PERSISTENT | ISEND | SEND | PREPARED | ACCUMULATE);
283 MPI_Request Request::recv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
285 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype,
286 src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE : comm->group()->index(src),
287 smpi_process()->index(), tag, comm, PERSISTENT | RECV | PREPARED);
290 MPI_Request Request::rma_recv_init(void *buf, int count, MPI_Datatype datatype, int src, int dst, int tag, MPI_Comm comm,
293 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
295 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
296 comm, RMA | NON_PERSISTENT | RECV | PREPARED);
298 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src, dst, tag,
299 comm, RMA | NON_PERSISTENT | RECV | PREPARED | ACCUMULATE);
305 MPI_Request Request::irecv_init(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
307 return new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
308 comm->group()->index(src), smpi_process()->index(), tag,
309 comm, PERSISTENT | RECV | PREPARED);
312 MPI_Request Request::isend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
314 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
315 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
316 comm->group()->index(dst), tag, comm, NON_PERSISTENT | ISEND | SEND);
321 MPI_Request Request::issend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
323 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
324 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
325 comm->group()->index(dst), tag,comm, NON_PERSISTENT | ISEND | SSEND | SEND);
331 MPI_Request Request::irecv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm)
333 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
334 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, src == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
335 comm->group()->index(src), smpi_process()->index(), tag, comm,
336 NON_PERSISTENT | RECV);
341 void Request::recv(void *buf, int count, MPI_Datatype datatype, int src, int tag, MPI_Comm comm, MPI_Status * status)
343 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
344 request = irecv(buf, count, datatype, src, tag, comm);
345 wait(&request,status);
349 void Request::send(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
351 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
352 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
353 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SEND);
356 wait(&request, MPI_STATUS_IGNORE);
360 void Request::ssend(void *buf, int count, MPI_Datatype datatype, int dst, int tag, MPI_Comm comm)
362 MPI_Request request = nullptr; /* MC needs the comm to be set to nullptr during the call */
363 request = new Request(buf==MPI_BOTTOM ? nullptr : buf, count, datatype, smpi_process()->index(),
364 comm->group()->index(dst), tag, comm, NON_PERSISTENT | SSEND | SEND);
367 wait(&request,MPI_STATUS_IGNORE);
371 void Request::sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype,int dst, int sendtag,
372 void *recvbuf, int recvcount, MPI_Datatype recvtype, int src, int recvtag,
373 MPI_Comm comm, MPI_Status * status)
375 MPI_Request requests[2];
377 int myid=smpi_process()->index();
378 if ((comm->group()->index(dst) == myid) && (comm->group()->index(src) == myid)){
379 Datatype::copy(sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype);
382 requests[0] = isend_init(sendbuf, sendcount, sendtype, dst, sendtag, comm);
383 requests[1] = irecv_init(recvbuf, recvcount, recvtype, src, recvtag, comm);
384 startall(2, requests);
385 waitall(2, requests, stats);
388 if(status != MPI_STATUS_IGNORE) {
389 // Copy receive status
394 void Request::start()
396 smx_mailbox_t mailbox;
398 xbt_assert(action_ == nullptr, "Cannot (re-)start unfinished communication");
403 if ((flags_ & RECV) != 0) {
404 this->print_request("New recv");
406 simgrid::smpi::Process* process = smpi_process_remote(dst_);
408 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
410 xbt_mutex_t mut = process->mailboxes_mutex();
411 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
412 xbt_mutex_acquire(mut);
414 if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
415 mailbox = process->mailbox();
417 else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
418 //We have to check both mailboxes (because SSEND messages are sent to the large mbox).
419 //begin with the more appropriate one : the small one.
420 mailbox = process->mailbox_small();
421 XBT_DEBUG("Is there a corresponding send already posted in the small mailbox %p (in case of SSEND)?", mailbox);
422 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv,
423 static_cast<void*>(this));
425 if (action == nullptr) {
426 mailbox = process->mailbox();
427 XBT_DEBUG("No, nothing in the small mailbox test the other one : %p", mailbox);
428 action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
429 if (action == nullptr) {
430 XBT_DEBUG("Still nothing, switch back to the small mailbox : %p", mailbox);
431 mailbox = process->mailbox_small();
434 XBT_DEBUG("yes there was something for us in the large mailbox");
437 mailbox = process->mailbox_small();
438 XBT_DEBUG("Is there a corresponding send already posted the small mailbox?");
439 smx_activity_t action = simcall_comm_iprobe(mailbox, 0, src_,tag_, &match_recv, static_cast<void*>(this));
441 if (action == nullptr) {
442 XBT_DEBUG("No, nothing in the permanent receive mailbox");
443 mailbox = process->mailbox();
445 XBT_DEBUG("yes there was something for us in the small mailbox");
449 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
451 action_ = simcall_comm_irecv(process->process(), mailbox, buf_, &real_size_, &match_recv,
452 ! process->replaying()? smpi_comm_copy_data_callback
453 : &smpi_comm_null_copy_buffer_callback, this, -1.0);
454 XBT_DEBUG("recv simcall posted");
456 if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
457 xbt_mutex_release(mut);
458 } else { /* the RECV flag was not set, so this is a send */
459 simgrid::smpi::Process* process = smpi_process_remote(dst_);
461 if (TRACE_smpi_view_internals()) {
462 TRACE_smpi_send(rank, rank, dst_, tag_, size_);
464 this->print_request("New send");
467 if ((flags_ & SSEND) == 0 && ( (flags_ & RMA) != 0
468 || static_cast<int>(size_) < xbt_cfg_get_int("smpi/send-is-detached-thresh") ) ) {
469 void *oldbuf = nullptr;
471 XBT_DEBUG("Send request %p is detached", this);
473 if(!(old_type_->flags() & DT_FLAG_DERIVED)){
475 if (!process->replaying() && oldbuf != nullptr && size_!=0){
476 if((smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP)
477 && (static_cast<char*>(buf_) >= smpi_start_data_exe)
478 && (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
479 XBT_DEBUG("Privatization : We are sending from a zone inside global memory. Switch data segment ");
480 smpi_switch_data_segment(src_);
482 buf = xbt_malloc(size_);
483 memcpy(buf,oldbuf,size_);
484 XBT_DEBUG("buf %p copied into %p",oldbuf,buf);
489 //if we are giving back the control to the user without waiting for completion, we have to inject timings
490 double sleeptime = 0.0;
491 if(detached_ != 0 || ((flags_ & (ISEND|SSEND)) != 0)){// issend should be treated as isend
492 //isend and send timings may be different
493 sleeptime = ((flags_ & ISEND) != 0) ? smpi_ois(size_) : smpi_os(size_);
497 simcall_process_sleep(sleeptime);
498 XBT_DEBUG("sending size of %zu : sleep %f ", size_, sleeptime);
501 int async_small_thresh = xbt_cfg_get_int("smpi/async-small-thresh");
503 xbt_mutex_t mut=process->mailboxes_mutex();
505 if (async_small_thresh != 0 || (flags_ & RMA) != 0)
506 xbt_mutex_acquire(mut);
508 if (!(async_small_thresh != 0 || (flags_ & RMA) !=0)) {
509 mailbox = process->mailbox();
510 } else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
511 mailbox = process->mailbox();
512 XBT_DEBUG("Is there a corresponding recv already posted in the large mailbox %p?", mailbox);
513 smx_activity_t action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send,
514 static_cast<void*>(this));
515 if (action == nullptr) {
516 if ((flags_ & SSEND) == 0){
517 mailbox = process->mailbox_small();
518 XBT_DEBUG("No, nothing in the large mailbox, message is to be sent on the small one %p", mailbox);
520 mailbox = process->mailbox_small();
521 XBT_DEBUG("SSEND : Is there a corresponding recv already posted in the small mailbox %p?", mailbox);
522 action = simcall_comm_iprobe(mailbox, 1,dst_, tag_, &match_send, static_cast<void*>(this));
523 if (action == nullptr) {
524 XBT_DEBUG("No, we are first, send to large mailbox");
525 mailbox = process->mailbox();
529 XBT_DEBUG("Yes there was something for us in the large mailbox");
532 mailbox = process->mailbox();
533 XBT_DEBUG("Send request %p is in the large mailbox %p (buf: %p)",mailbox, this,buf_);
536 // we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
538 action_ = simcall_comm_isend(SIMIX_process_from_PID(src_+1), mailbox, size_, -1.0,
539 buf, real_size_, &match_send,
540 &xbt_free_f, // how to free the userdata if a detached send fails
541 !process->replaying() ? smpi_comm_copy_data_callback
542 : &smpi_comm_null_copy_buffer_callback, this,
543 // detach if msg size < eager/rdv switch limit
545 XBT_DEBUG("send simcall posted");
547 /* FIXME: detached sends are not traceable (action_ == nullptr) */
548 if (action_ != nullptr)
549 simcall_set_category(action_, TRACE_internal_smpi_get_category());
550 if (async_small_thresh != 0 || ((flags_ & RMA)!=0))
551 xbt_mutex_release(mut);
555 void Request::startall(int count, MPI_Request * requests)
557 if(requests== nullptr)
560 for(int i = 0; i < count; i++) {
561 requests[i]->start();
565 int Request::test(MPI_Request * request, MPI_Status * status) {
566 //assume that request is not MPI_REQUEST_NULL (filtered in PMPI_Test or testall before)
567 // to avoid deadlocks if used as a break condition, such as
568 // while (MPI_Test(request, flag, status) && flag) dostuff...
569 // because the time will not normally advance when only calls to MPI_Test are made -> deadlock
570 // multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
571 static int nsleeps = 1;
572 if(smpi_test_sleep > 0)
573 simcall_process_sleep(nsleeps*smpi_test_sleep);
575 Status::empty(status);
577 if (((*request)->flags_ & PREPARED) == 0) {
578 if ((*request)->action_ != nullptr)
579 flag = simcall_comm_test((*request)->action_);
581 finish_wait(request,status);
582 nsleeps=1;//reset the number of sleeps we will do next time
583 if (*request != MPI_REQUEST_NULL && ((*request)->flags_ & PERSISTENT)==0)
584 *request = MPI_REQUEST_NULL;
585 } else if (xbt_cfg_get_boolean("smpi/grow-injected-times")){
592 int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
598 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
600 for(i = 0; i < incount; i++) {
601 if((requests[i] != MPI_REQUEST_NULL)) {
602 if(test(&requests[i], pstat)) {
605 if(status != MPI_STATUSES_IGNORE) {
608 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags_ & NON_PERSISTENT)
609 requests[i]=MPI_REQUEST_NULL;
615 if(count_dead==incount)
616 return MPI_UNDEFINED;
620 int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
622 std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
623 comms.reserve(count);
628 *index = MPI_UNDEFINED;
630 std::vector<int> map; /** Maps all matching comms back to their location in requests **/
631 for(i = 0; i < count; i++) {
632 if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && !(requests[i]->flags_ & PREPARED)) {
633 comms.push_back(requests[i]->action_);
638 //multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
639 static int nsleeps = 1;
640 if(smpi_test_sleep > 0)
641 simcall_process_sleep(nsleeps*smpi_test_sleep);
643 i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
644 if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
646 finish_wait(&requests[*index],status);
649 if (requests[*index] != MPI_REQUEST_NULL && (requests[*index]->flags_ & NON_PERSISTENT)) {
650 requests[*index] = MPI_REQUEST_NULL;
656 //all requests are null or inactive, return true
658 Status::empty(status);
664 int Request::testall(int count, MPI_Request requests[], MPI_Status status[])
667 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
669 for(int i=0; i<count; i++){
670 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED)) {
671 if (test(&requests[i], pstat)!=1){
674 requests[i]=MPI_REQUEST_NULL;
677 Status::empty(pstat);
679 if(status != MPI_STATUSES_IGNORE) {
686 void Request::probe(int source, int tag, MPI_Comm comm, MPI_Status* status){
688 //FIXME find another way to avoid busy waiting ?
689 // the issue here is that we have to wait on a nonexistent comm
691 iprobe(source, tag, comm, &flag, status);
692 XBT_DEBUG("Busy Waiting on probing : %d", flag);
696 void Request::iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* status){
697 // to avoid deadlock, we have to sleep some time here, or the timer won't advance and we will only do iprobe simcalls
698 // especially when used as a break condition, such as while (MPI_Iprobe(...)) dostuff...
699 // nsleeps is a multiplier to the sleeptime, to increase speed of execution, each failed iprobe will increase it
700 // This can speed up the execution of certain applications by an order of magnitude, such as HPL
701 static int nsleeps = 1;
702 double speed = simgrid::s4u::Actor::self()->host()->speed();
703 double maxrate = xbt_cfg_get_double("smpi/iprobe-cpu-usage");
704 MPI_Request request = new Request(nullptr, 0, MPI_CHAR, source == MPI_ANY_SOURCE ? MPI_ANY_SOURCE :
705 comm->group()->index(source), comm->rank(), tag, comm, PERSISTENT | RECV);
706 if (smpi_iprobe_sleep > 0) {
707 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);
708 simcall_execution_wait(iprobe_sleep);
710 // behave like a receive, but don't do it
711 smx_mailbox_t mailbox;
713 request->print_request("New iprobe");
714 // We have to test both mailboxes as we don't know if we will receive one one or another
715 if (xbt_cfg_get_int("smpi/async-small-thresh") > 0){
716 mailbox = smpi_process()->mailbox_small();
717 XBT_DEBUG("Trying to probe the perm recv mailbox");
718 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_, request->tag_, &match_recv,
719 static_cast<void*>(request));
722 if (request->action_ == nullptr){
723 mailbox = smpi_process()->mailbox();
724 XBT_DEBUG("trying to probe the other mailbox");
725 request->action_ = simcall_comm_iprobe(mailbox, 0, request->src_,request->tag_, &match_recv,
726 static_cast<void*>(request));
729 if (request->action_ != nullptr){
730 simgrid::kernel::activity::Comm *sync_comm = static_cast<simgrid::kernel::activity::Comm*>(request->action_);
731 MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
733 if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
734 status->MPI_SOURCE = comm->group()->rank(req->src_);
735 status->MPI_TAG = req->tag_;
736 status->MPI_ERROR = MPI_SUCCESS;
737 status->count = req->real_size_;
739 nsleeps = 1;//reset the number of sleeps we will do next time
743 if (xbt_cfg_get_boolean("smpi/grow-injected-times"))
749 void Request::finish_wait(MPI_Request* request, MPI_Status * status)
751 MPI_Request req = *request;
752 Status::empty(status);
754 if(!((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)){
755 if(status != MPI_STATUS_IGNORE) {
756 int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
757 status->MPI_SOURCE = req->comm_->group()->rank(src);
758 status->MPI_TAG = req->tag_ == MPI_ANY_TAG ? req->real_tag_ : req->tag_;
759 status->MPI_ERROR = req->truncated_ != 0 ? MPI_ERR_TRUNCATE : MPI_SUCCESS;
760 // this handles the case were size in receive differs from size in send
761 status->count = req->real_size_;
764 req->print_request("Finishing");
765 MPI_Datatype datatype = req->old_type_;
767 // FIXME Handle the case of a partial shared malloc.
768 if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED))){// && (!smpi_is_shared(req->old_buf_))){
770 if (!smpi_process()->replaying()){
771 if( smpi_privatize_global_variables == SMPI_PRIVATIZE_MMAP && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
772 && ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
773 XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
774 smpi_switch_data_segment(smpi_process()->index());
778 if(datatype->flags() & DT_FLAG_DERIVED){
779 // This part handles the problem of non-contignous memory the unserialization at the reception
780 if((req->flags_ & RECV) && datatype->size()!=0)
781 datatype->unserialize(req->buf_, req->old_buf_, req->real_size_/datatype->size() , req->op_);
783 }else if(req->flags_ & RECV){//apply op on contiguous buffer for accumulate
784 if(datatype->size()!=0){
785 int n =req->real_size_/datatype->size();
786 req->op_->apply(req->buf_, req->old_buf_, &n, datatype);
793 if (TRACE_smpi_view_internals() && ((req->flags_ & RECV) != 0)){
794 int rank = smpi_process()->index();
795 int src_traced = (req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_);
796 TRACE_smpi_recv(rank, src_traced, rank,req->tag_);
798 if(req->detached_sender_ != nullptr){
799 //integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
800 double sleeptime = smpi_or(req->real_size_);
802 simcall_process_sleep(sleeptime);
803 XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
805 unref(&(req->detached_sender_));
807 if(req->flags_ & PERSISTENT)
808 req->action_ = nullptr;
809 req->flags_ |= FINISHED;
813 void Request::wait(MPI_Request * request, MPI_Status * status)
815 (*request)->print_request("Waiting");
816 if ((*request)->flags_ & PREPARED) {
817 Status::empty(status);
821 if ((*request)->action_ != nullptr)
822 // this is not a detached send
823 simcall_comm_wait((*request)->action_, -1.0);
825 finish_wait(request,status);
826 if (*request != MPI_REQUEST_NULL && (((*request)->flags_ & NON_PERSISTENT)!=0))
827 *request = MPI_REQUEST_NULL;
830 int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
832 s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
835 int index = MPI_UNDEFINED;
839 // Wait for a request to complete
840 xbt_dynar_init(&comms, sizeof(smx_activity_t), nullptr);
841 map = xbt_new(int, count);
842 XBT_DEBUG("Wait for one of %d", count);
843 for(i = 0; i < count; i++) {
844 if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED) && !(requests[i]->flags_ & FINISHED)) {
845 if (requests[i]->action_ != nullptr) {
846 XBT_DEBUG("Waiting any %p ", requests[i]);
847 xbt_dynar_push(&comms, &requests[i]->action_);
851 // This is a finished detached request, let's return this one
852 size = 0; // so we free the dynar but don't do the waitany call
854 finish_wait(&requests[i], status); // cleanup if refcount = 0
855 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
856 requests[i] = MPI_REQUEST_NULL; // set to null
862 i = simcall_comm_waitany(&comms, -1);
864 // not MPI_UNDEFINED, as this is a simix return code
867 //in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
868 if ((requests[index] == MPI_REQUEST_NULL)
869 || (!((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))){
870 finish_wait(&requests[index],status);
871 if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
872 requests[index] = MPI_REQUEST_NULL;
877 xbt_dynar_free_data(&comms);
881 if (index==MPI_UNDEFINED)
882 Status::empty(status);
887 static int sort_accumulates(MPI_Request a, MPI_Request b)
889 return (a->tag() > b->tag());
892 int Request::waitall(int count, MPI_Request requests[], MPI_Status status[])
894 std::vector<MPI_Request> accumulates;
897 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
898 int retvalue = MPI_SUCCESS;
899 //tag invalid requests in the set
900 if (status != MPI_STATUSES_IGNORE) {
901 for (int c = 0; c < count; c++) {
902 if (requests[c] == MPI_REQUEST_NULL || requests[c]->dst_ == MPI_PROC_NULL || (requests[c]->flags_ & PREPARED)) {
903 Status::empty(&status[c]);
904 } else if (requests[c]->src_ == MPI_PROC_NULL) {
905 Status::empty(&status[c]);
906 status[c].MPI_SOURCE = MPI_PROC_NULL;
910 for (int c = 0; c < count; c++) {
911 if (MC_is_active() || MC_record_replay_is_active()) {
912 wait(&requests[c],pstat);
915 index = waitany(count, requests, pstat);
916 if (index == MPI_UNDEFINED)
919 if (requests[index] != MPI_REQUEST_NULL
920 && (requests[index]->flags_ & RECV)
921 && (requests[index]->flags_ & ACCUMULATE))
922 accumulates.push_back(requests[index]);
923 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
924 requests[index] = MPI_REQUEST_NULL;
926 if (status != MPI_STATUSES_IGNORE) {
927 status[index] = *pstat;
928 if (status[index].MPI_ERROR == MPI_ERR_TRUNCATE)
929 retvalue = MPI_ERR_IN_STATUS;
933 if (!accumulates.empty()) {
934 std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
935 for (auto req : accumulates) {
936 finish_wait(&req, status);
943 int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
949 MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
951 for(i = 0; i < incount; i++)
953 index=waitany(incount, requests, pstat);
954 if(index!=MPI_UNDEFINED){
955 indices[count] = index;
957 if(status != MPI_STATUSES_IGNORE) {
958 status[index] = *pstat;
960 if (requests[index] != MPI_REQUEST_NULL && (requests[index]->flags_ & NON_PERSISTENT))
961 requests[index]=MPI_REQUEST_NULL;
963 return MPI_UNDEFINED;
969 MPI_Request Request::f2c(int id) {
971 if(id==MPI_FORTRAN_REQUEST_NULL)
972 return static_cast<MPI_Request>(MPI_REQUEST_NULL);
973 return static_cast<MPI_Request>(xbt_dict_get(F2C::f2c_lookup(), get_key_id(key, id)));
976 int Request::add_f() {
977 if(F2C::f2c_lookup()==nullptr){
978 F2C::set_f2c_lookup(xbt_dict_new_homogeneous(nullptr));
981 xbt_dict_set(F2C::f2c_lookup(), get_key_id(key, F2C::f2c_id()), this, nullptr);
982 F2C::f2c_id_increment();
983 return F2C::f2c_id()-1;
986 void Request::free_f(int id) {
988 if(id!=MPI_FORTRAN_REQUEST_NULL)
989 xbt_dict_remove(F2C::f2c_lookup(), get_key_id(key, id));