/* This program is free software; you can redistribute it and/or modify it
* under the terms of the license (GNU LGPL) which comes with this package. */
-#include <xbt/config.hpp>
-#include <algorithm>
-#include "private.h"
+#include "src/smpi/smpi_request.hpp"
+
#include "mc/mc.h"
+#include "src/kernel/activity/CommImpl.hpp"
#include "src/mc/mc_replay.h"
-#include "src/simix/smx_private.h"
-#include "simgrid/sg_config.h"
-#include "smpi/smpi_utils.hpp"
-#include <simgrid/s4u/host.hpp>
-#include "src/kernel/activity/SynchroComm.hpp"
+#include "src/smpi/SmpiHost.hpp"
+#include "src/smpi/private.h"
+#include "src/smpi/smpi_comm.hpp"
+#include "src/smpi/smpi_datatype.hpp"
+#include "src/smpi/smpi_op.hpp"
+#include "src/smpi/smpi_process.hpp"
+
+#include <algorithm>
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_request, smpi, "Logging specific to SMPI (reques)");
static simgrid::config::Flag<double> smpi_test_sleep(
"smpi/test", "Minimum time to inject inside a call to MPI_Test", 1e-4);
-std::vector<s_smpi_factor_t> smpi_os_values;
-std::vector<s_smpi_factor_t> smpi_or_values;
std::vector<s_smpi_factor_t> smpi_ois_values;
extern void (*smpi_comm_copy_data_callback) (smx_activity_t, void*, size_t);
-static double smpi_os(size_t size)
-{
- if (smpi_os_values.empty()) {
- smpi_os_values = parse_factor(xbt_cfg_get_string("smpi/os"));
- }
- double current=smpi_os_values.empty()?0.0:smpi_os_values[0].values[0]+smpi_os_values[0].values[1]*size;
- // Iterate over all the sections that were specified and find the right
- // value. (fact.factor represents the interval sizes; we want to find the
- // section that has fact.factor <= size and no other such fact.factor <= size)
- // Note: parse_factor() (used before) already sorts the vector we iterate over!
- for (auto& fact : smpi_os_values) {
- if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
- XBT_DEBUG("os : %zu <= %zu return %.10f", size, fact.factor, current);
- return current;
- }else{
- // If the next section is too large, the current section must be used.
- // Hence, save the cost, as we might have to use it.
- current = fact.values[0]+fact.values[1]*size;
- }
- }
- XBT_DEBUG("Searching for smpi/os: %zu is larger than the largest boundary, return %.10f", size, current);
-
- return current;
-}
-
-static double smpi_ois(size_t size)
-{
- if (smpi_ois_values.empty()) {
- smpi_ois_values = parse_factor(xbt_cfg_get_string("smpi/ois"));
- }
- double current=smpi_ois_values.empty()?0.0:smpi_ois_values[0].values[0]+smpi_ois_values[0].values[1]*size;
- // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
- // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
- // Note: parse_factor() (used before) already sorts the vector we iterate over!
- for (auto& fact : smpi_ois_values) {
- if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
- XBT_DEBUG("ois : %zu <= %zu return %.10f", size, fact.factor, current);
- return current;
- }else{
- // If the next section is too large, the current section must be used.
- // Hence, save the cost, as we might have to use it.
- current = fact.values[0]+fact.values[1]*size;
- }
- }
- XBT_DEBUG("Searching for smpi/ois: %zu is larger than the largest boundary, return %.10f", size, current);
-
- return current;
-}
-
-static double smpi_or(size_t size)
-{
- if (smpi_or_values.empty()) {
- smpi_or_values = parse_factor(xbt_cfg_get_string("smpi/or"));
- }
-
- double current=smpi_or_values.empty()?0.0:smpi_or_values.front().values[0]+smpi_or_values.front().values[1]*size;
-
- // Iterate over all the sections that were specified and find the right value. (fact.factor represents the interval
- // sizes; we want to find the section that has fact.factor <= size and no other such fact.factor <= size)
- // Note: parse_factor() (used before) already sorts the vector we iterate over!
- for (auto fact : smpi_or_values) {
- if (size <= fact.factor) { // Values already too large, use the previously computed value of current!
- XBT_DEBUG("or : %zu <= %zu return %.10f", size, fact.factor, current);
- return current;
- } else {
- // If the next section is too large, the current section must be used.
- // Hence, save the cost, as we might have to use it.
- current=fact.values[0]+fact.values[1]*size;
- }
- }
- XBT_DEBUG("smpi_or: %zu is larger than largest boundary, return %.10f", size, current);
-
- return current;
-}
-
namespace simgrid{
namespace smpi{
-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)
+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)
{
void *old_buf = nullptr;
- if(((((flags & RECV) != 0) && ((flags & ACCUMULATE) !=0)) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(buf_))){
+// FIXME Handle the case of a partial shared malloc.
+ if ((((flags & RECV) != 0) && ((flags & ACCUMULATE) != 0)) || (datatype->flags() & DT_FLAG_DERIVED)) {
// This part handles the problem of non-contiguous memory
old_buf = buf;
- buf_ = count==0 ? nullptr : xbt_malloc(count*datatype->size());
- if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
- datatype->serialize(old_buf, buf_, count);
+ if (count==0){
+ buf_ = nullptr;
+ }else {
+ buf_ = xbt_malloc(count*datatype->size());
+ if ((datatype->flags() & DT_FLAG_DERIVED) && ((flags & SEND) != 0)) {
+ datatype->serialize(old_buf, buf_, count);
+ }
}
}
// This part handles the problem of non-contiguous memory (for the unserialisation at the reception)
}
}
-int Request::match_recv(void* a, void* b, smx_activity_t ignored) {
+int Request::match_recv(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
+{
MPI_Request ref = static_cast<MPI_Request>(a);
MPI_Request req = static_cast<MPI_Request>(b);
XBT_DEBUG("Trying to match a recv of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
ref->real_src_ = req->src_;
if(ref->tag_ == MPI_ANY_TAG)
ref->real_tag_ = req->tag_;
- if(ref->real_size_ < req->real_size_)
+ if(ref->real_size_ < req->real_size_)
ref->truncated_ = 1;
if(req->detached_==1)
ref->detached_sender_=req; //tie the sender to the receiver, as it is detached and has to be freed in the receiver
}else return 0;
}
-int Request::match_send(void* a, void* b,smx_activity_t ignored) {
+int Request::match_send(void* a, void* b, simgrid::kernel::activity::CommImpl* ignored)
+{
MPI_Request ref = static_cast<MPI_Request>(a);
MPI_Request req = static_cast<MPI_Request>(b);
XBT_DEBUG("Trying to match a send of src %d against %d, tag %d against %d",ref->src_,req->src_, ref->tag_, req->tag_);
if (async_small_thresh == 0 && (flags_ & RMA) == 0 ) {
mailbox = process->mailbox();
- }
+ }
else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) {
//We have to check both mailboxes (because SSEND messages are sent to the large mbox).
//begin with the more appropriate one : the small one.
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
real_size_=size_;
- action_ = simcall_comm_irecv(process->process(), mailbox, buf_, &real_size_, &match_recv,
- ! process->replaying()? smpi_comm_copy_data_callback
- : &smpi_comm_null_copy_buffer_callback, this, -1.0);
+ action_ = simcall_comm_irecv(
+ process->process(), mailbox, buf_, &real_size_, &match_recv,
+ process->replaying() ? &smpi_comm_null_copy_buffer_callback : smpi_comm_copy_data_callback, this, -1.0);
XBT_DEBUG("recv simcall posted");
if (async_small_thresh != 0 || (flags_ & RMA) != 0 )
detached_ = 1;
XBT_DEBUG("Send request %p is detached", this);
refcount_++;
- if(!(old_type_->flags() & DT_FLAG_DERIVED)){
+ if (not(old_type_->flags() & DT_FLAG_DERIVED)) {
oldbuf = buf_;
- if (!process->replaying() && oldbuf != nullptr && size_!=0){
+ if (not process->replaying() && oldbuf != nullptr && size_ != 0) {
if((smpi_privatize_global_variables != 0)
&& (static_cast<char*>(buf_) >= smpi_start_data_exe)
&& (static_cast<char*>(buf_) < smpi_start_data_exe + smpi_size_data_exe )){
//if we are giving back the control to the user without waiting for completion, we have to inject timings
double sleeptime = 0.0;
- if(detached_ != 0 || ((flags_ & (ISEND|SSEND)) != 0)){// issend should be treated as isend
- //isend and send timings may be different
- sleeptime = ((flags_ & ISEND) != 0) ? smpi_ois(size_) : smpi_os(size_);
+ if (detached_ != 0 || ((flags_ & (ISEND | SSEND)) != 0)) { // issend should be treated as isend
+ // isend and send timings may be different
+ sleeptime = ((flags_ & ISEND) != 0)
+ ? simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->oisend(size_)
+ : simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->osend(size_);
}
if(sleeptime > 0.0){
if (async_small_thresh != 0 || (flags_ & RMA) != 0)
xbt_mutex_acquire(mut);
- if (!(async_small_thresh != 0 || (flags_ & RMA) !=0)) {
+ if (not(async_small_thresh != 0 || (flags_ & RMA) != 0)) {
mailbox = process->mailbox();
} else if (((flags_ & RMA) != 0) || static_cast<int>(size_) < async_small_thresh) { // eager mode
mailbox = process->mailbox();
// we make a copy here, as the size is modified by simix, and we may reuse the request in another receive later
real_size_=size_;
- action_ = simcall_comm_isend(SIMIX_process_from_PID(src_+1), mailbox, size_, -1.0,
- buf, real_size_, &match_send,
- &xbt_free_f, // how to free the userdata if a detached send fails
- !process->replaying() ? smpi_comm_copy_data_callback
- : &smpi_comm_null_copy_buffer_callback, this,
- // detach if msg size < eager/rdv switch limit
- detached_);
+ action_ = simcall_comm_isend(
+ SIMIX_process_from_PID(src_ + 1), mailbox, size_, -1.0, buf, real_size_, &match_send,
+ &xbt_free_f, // how to free the userdata if a detached send fails
+ not process->replaying() ? smpi_comm_copy_data_callback : &smpi_comm_null_copy_buffer_callback, this,
+ // detach if msg size < eager/rdv switch limit
+ detached_);
XBT_DEBUG("send simcall posted");
/* FIXME: detached sends are not traceable (action_ == nullptr) */
void Request::startall(int count, MPI_Request * requests)
{
- if(requests== nullptr)
+ if(requests== nullptr)
return;
for(int i = 0; i < count; i++) {
// because the time will not normally advance when only calls to MPI_Test are made -> deadlock
// multiplier to the sleeptime, to increase speed of execution, each failed test will increase it
static int nsleeps = 1;
- if(smpi_test_sleep > 0)
+ if(smpi_test_sleep > 0)
simcall_process_sleep(nsleeps*smpi_test_sleep);
Status::empty(status);
int Request::testsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
{
- int i;
int count = 0;
int count_dead = 0;
MPI_Status stat;
MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
- for(i = 0; i < incount; i++) {
- if((requests[i] != MPI_REQUEST_NULL)) {
- if(test(&requests[i], pstat)) {
- indices[i] = 1;
- count++;
- if(status != MPI_STATUSES_IGNORE) {
- status[i] = *pstat;
- }
- if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->flags_ & NON_PERSISTENT)
- requests[i]=MPI_REQUEST_NULL;
+ for (int i = 0; i < incount; i++) {
+ if (requests[i] != MPI_REQUEST_NULL) {
+ if (test(&requests[i], pstat)) {
+ indices[i] = 1;
+ count++;
+ if (status != MPI_STATUSES_IGNORE)
+ status[i] = *pstat;
+ if ((requests[i] != MPI_REQUEST_NULL) && (requests[i]->flags_ & NON_PERSISTENT))
+ requests[i] = MPI_REQUEST_NULL;
}
- }else{
+ } else {
count_dead++;
}
}
int Request::testany(int count, MPI_Request requests[], int *index, MPI_Status * status)
{
- std::vector<simgrid::kernel::activity::ActivityImpl*> comms;
+ std::vector<simgrid::kernel::activity::ActivityImplPtr> comms;
comms.reserve(count);
int i;
std::vector<int> map; /** Maps all matching comms back to their location in requests **/
for(i = 0; i < count; i++) {
- if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && !(requests[i]->flags_ & PREPARED)) {
- comms.push_back(requests[i]->action_);
- map.push_back(i);
+ if ((requests[i] != MPI_REQUEST_NULL) && requests[i]->action_ && not(requests[i]->flags_ & PREPARED)) {
+ comms.push_back(requests[i]->action_);
+ map.push_back(i);
}
}
- if(!map.empty()) {
+ if (not map.empty()) {
//multiplier to the sleeptime, to increase speed of execution, each failed testany will increase it
static int nsleeps = 1;
- if(smpi_test_sleep > 0)
+ if(smpi_test_sleep > 0)
simcall_process_sleep(nsleeps*smpi_test_sleep);
i = simcall_comm_testany(comms.data(), comms.size()); // The i-th element in comms matches!
if (i != -1) { // -1 is not MPI_UNDEFINED but a SIMIX return code. (nothing matches)
- *index = map[i];
+ *index = map[i];
finish_wait(&requests[*index],status);
flag = 1;
nsleeps = 1;
MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
int flag=1;
for(int i=0; i<count; i++){
- if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED)) {
+ if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED)) {
if (test(&requests[i], pstat)!=1){
flag=0;
}else{
}
if (request->action_ != nullptr){
- simgrid::kernel::activity::Comm *sync_comm = static_cast<simgrid::kernel::activity::Comm*>(request->action_);
+ simgrid::kernel::activity::CommImplPtr sync_comm =
+ boost::static_pointer_cast<simgrid::kernel::activity::CommImpl>(request->action_);
MPI_Request req = static_cast<MPI_Request>(sync_comm->src_data);
*flag = 1;
if(status != MPI_STATUS_IGNORE && (req->flags_ & PREPARED) == 0) {
MPI_Request req = *request;
Status::empty(status);
- if(!((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)){
+ if (not((req->detached_ != 0) && ((req->flags_ & SEND) != 0)) && ((req->flags_ & PREPARED) == 0)) {
if(status != MPI_STATUS_IGNORE) {
int src = req->src_ == MPI_ANY_SOURCE ? req->real_src_ : req->src_;
status->MPI_SOURCE = req->comm_->group()->rank(src);
req->print_request("Finishing");
MPI_Datatype datatype = req->old_type_;
- if((((req->flags_ & ACCUMULATE) != 0) || (datatype->flags() & DT_FLAG_DERIVED)) && (!smpi_is_shared(req->old_buf_))){
+// FIXME Handle the case of a partial shared malloc.
+ if (((req->flags_ & ACCUMULATE) != 0) ||
+ (datatype->flags() & DT_FLAG_DERIVED)) { // && (not smpi_is_shared(req->old_buf_))){
- if (!smpi_process()->replaying()){
+ if (not smpi_process()->replaying()) {
if( smpi_privatize_global_variables != 0 && (static_cast<char*>(req->old_buf_) >= smpi_start_data_exe)
&& ((char*)req->old_buf_ < smpi_start_data_exe + smpi_size_data_exe )){
XBT_VERB("Privatization : We are unserializing to a zone in global memory Switch data segment ");
}
if(req->detached_sender_ != nullptr){
//integrate pseudo-timing for buffering of small messages, do not bother to execute the simcall if 0
- double sleeptime = smpi_or(req->real_size_);
+ double sleeptime = simgrid::s4u::Actor::self()->host()->extension<simgrid::smpi::SmpiHost>()->orecv(req->real_size());
if(sleeptime > 0.0){
simcall_process_sleep(sleeptime);
XBT_DEBUG("receiving size of %zu : sleep %f ", req->real_size_, sleeptime);
int Request::waitany(int count, MPI_Request requests[], MPI_Status * status)
{
s_xbt_dynar_t comms; // Keep it on stack to save some extra mallocs
- int i;
- int size = 0;
int index = MPI_UNDEFINED;
- int *map;
if(count > 0) {
+ int size = 0;
// Wait for a request to complete
- xbt_dynar_init(&comms, sizeof(smx_activity_t), nullptr);
- map = xbt_new(int, count);
+ xbt_dynar_init(&comms, sizeof(smx_activity_t), [](void*ptr){
+ intrusive_ptr_release(*(simgrid::kernel::activity::ActivityImpl**)ptr);
+ });
+ int *map = xbt_new(int, count);
XBT_DEBUG("Wait for one of %d", count);
- for(i = 0; i < count; i++) {
- if (requests[i] != MPI_REQUEST_NULL && !(requests[i]->flags_ & PREPARED) && !(requests[i]->flags_ & FINISHED)) {
+ for(int i = 0; i < count; i++) {
+ if (requests[i] != MPI_REQUEST_NULL && not(requests[i]->flags_ & PREPARED) &&
+ not(requests[i]->flags_ & FINISHED)) {
if (requests[i]->action_ != nullptr) {
XBT_DEBUG("Waiting any %p ", requests[i]);
- xbt_dynar_push(&comms, &requests[i]->action_);
+ intrusive_ptr_add_ref(requests[i]->action_.get());
+ xbt_dynar_push_as(&comms, simgrid::kernel::activity::ActivityImpl*, requests[i]->action_.get());
map[size] = i;
size++;
} else {
}
}
}
- if(size > 0) {
- i = simcall_comm_waitany(&comms, -1);
+ if (size > 0) {
+ XBT_DEBUG("Enter waitany for %lu comms", xbt_dynar_length(&comms));
+ int i = simcall_comm_waitany(&comms, -1);
// not MPI_UNDEFINED, as this is a simix return code
if (i != -1) {
index = map[i];
//in case of an accumulate, we have to wait the end of all requests to apply the operation, ordered correctly.
- if ((requests[index] == MPI_REQUEST_NULL)
- || (!((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))){
+ if ((requests[index] == MPI_REQUEST_NULL) ||
+ (not((requests[index]->flags_ & ACCUMULATE) && (requests[index]->flags_ & RECV)))) {
finish_wait(&requests[index],status);
if (requests[i] != MPI_REQUEST_NULL && (requests[i]->flags_ & NON_PERSISTENT))
requests[index] = MPI_REQUEST_NULL;
std::vector<MPI_Request> accumulates;
int index;
MPI_Status stat;
- MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
+ MPI_Status *pstat = (status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat);
int retvalue = MPI_SUCCESS;
//tag invalid requests in the set
if (status != MPI_STATUSES_IGNORE) {
wait(&requests[c],pstat);
index = c;
} else {
- index = waitany(count, requests, pstat);
+ index = waitany(count, (MPI_Request*)requests, pstat);
if (index == MPI_UNDEFINED)
break;
}
}
- if (!accumulates.empty()) {
+ if (not accumulates.empty()) {
std::sort(accumulates.begin(), accumulates.end(), sort_accumulates);
for (auto req : accumulates) {
finish_wait(&req, status);
int Request::waitsome(int incount, MPI_Request requests[], int *indices, MPI_Status status[])
{
- int i;
int count = 0;
- int index;
MPI_Status stat;
MPI_Status *pstat = status == MPI_STATUSES_IGNORE ? MPI_STATUS_IGNORE : &stat;
- for(i = 0; i < incount; i++)
- {
- index=waitany(incount, requests, pstat);
+ for (int i = 0; i < incount; i++) {
+ int index = waitany(incount, requests, pstat);
if(index!=MPI_UNDEFINED){
indices[count] = index;
count++;
}
void Request::free_f(int id) {
- char key[KEY_SIZE];
- if(id!=MPI_FORTRAN_REQUEST_NULL)
+ if (id != MPI_FORTRAN_REQUEST_NULL) {
+ char key[KEY_SIZE];
xbt_dict_remove(F2C::f2c_lookup(), get_key_id(key, id));
+ }
}
}