-/* Copyright (c) 2009-2021. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2009-2022. The SimGrid Team. All rights reserved. */
/* 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 "smpi_replay.hpp"
+#include "simgrid/s4u/Exec.hpp"
#include "smpi_coll.hpp"
#include "smpi_comm.hpp"
+#include "smpi_config.hpp"
#include "smpi_datatype.hpp"
#include "smpi_group.hpp"
#include "smpi_request.hpp"
-#include "smpi_config.hpp"
-#include "simgrid/s4u/Exec.hpp"
+#include "src/smpi/include/private.hpp"
#include "xbt/replay.hpp"
-#include <simgrid/smpi/smpi_replay.hpp>
-#include <src/smpi/include/private.hpp>
+#include "xbt/str.h"
#include <cmath>
#include <limits>
#include <vector>
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay, smpi, "Trace Replay with SMPI");
-
// From https://stackoverflow.com/questions/7110301/generic-hash-for-tuples-in-unordered-map-unordered-set
// This is all just to make std::unordered_map work with std::tuple. If we need this in other places,
// this could go into a header file.
return i < action.size() ? simgrid::smpi::Datatype::decode(action[i]) : simgrid::smpi::replay::MPI_DEFAULT_TYPE;
}
-namespace simgrid {
-namespace smpi {
-
-namespace replay {
+namespace simgrid::smpi::replay {
MPI_Datatype MPI_DEFAULT_TYPE;
class RequestStorage {
private:
using req_key_t = std::tuple</*sender*/ int, /* receiver */ int, /* tag */ int>;
- using req_storage_t = std::unordered_map<req_key_t, MPI_Request, hash_tuple::hash<std::tuple<int, int, int>>>;
+ using req_storage_t = std::unordered_map<req_key_t, std::list<MPI_Request>, hash_tuple::hash<std::tuple<int, int, int>>>;
req_storage_t store;
void get_requests(std::vector<MPI_Request>& vec) const
{
- for (auto const& pair : store) {
- auto& req = pair.second;
+ for (auto const& [_, reqs] : store) {
aid_t my_proc_id = simgrid::s4u::this_actor::get_pid();
- if (req != MPI_REQUEST_NULL && (req->src() == my_proc_id || req->dst() == my_proc_id)) {
- vec.push_back(pair.second);
- pair.second->print_request("MM");
+ for (auto& req: reqs){
+ if (req != MPI_REQUEST_NULL && (req->src() == my_proc_id || req->dst() == my_proc_id)) {
+ vec.push_back(req);
+ req->print_request("MM");
+ }
}
}
}
- MPI_Request find(int src, int dst, int tag)
- {
- auto it = store.find(req_key_t(src, dst, tag));
- return (it == store.end()) ? MPI_REQUEST_NULL : it->second;
- }
-
- void remove(const Request* req)
- {
- if (req == MPI_REQUEST_NULL) return;
-
- store.erase(req_key_t(req->src()-1, req->dst()-1, req->tag()));
- }
+ MPI_Request pop(int src, int dst, int tag)
+ {
+ auto it = store.find(req_key_t(src, dst, tag));
+ if (it == store.end())
+ return MPI_REQUEST_NULL;
+ MPI_Request req = it->second.front();
+ it->second.pop_front();
+ if(it->second.empty())
+ store.erase(req_key_t(src, dst, tag));
+ return req;
+ }
- void add(MPI_Request req)
- {
- if (req != MPI_REQUEST_NULL) // Can and does happen in the case of TestAction
- store.insert({req_key_t(req->src()-1, req->dst()-1, req->tag()), req});
+ void add(MPI_Request req)
+ {
+ if (req != MPI_REQUEST_NULL){ // Can and does happen in the case of TestAction
+ store[req_key_t(req->src()-1, req->dst()-1, req->tag())].push_back(req);
}
+ }
- /* Sometimes we need to re-insert MPI_REQUEST_NULL but we still need src,dst and tag */
- void addNullRequest(int src, int dst, int tag)
- {
- store.insert({req_key_t(MPI_COMM_WORLD->group()->actor(src) - 1, MPI_COMM_WORLD->group()->actor(dst) - 1, tag),
- MPI_REQUEST_NULL});
- }
+ /* Sometimes we need to re-insert MPI_REQUEST_NULL but we still need src,dst and tag */
+ void addNullRequest(int src, int dst, int tag)
+ {
+ int src_pid = MPI_COMM_WORLD->group()->actor(src) - 1;
+ int dest_pid = MPI_COMM_WORLD->group()->actor(dst) - 1;
+ store[req_key_t(src_pid, dest_pid, tag)].push_back(MPI_REQUEST_NULL);
+ }
};
void WaitTestParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
std::string s = boost::algorithm::join(action, " ");
xbt_assert(req_storage.size(), "action wait not preceded by any irecv or isend: %s", s.c_str());
const WaitTestParser& args = get_args();
- MPI_Request request = req_storage.find(args.src, args.dst, args.tag);
- req_storage.remove(request);
+ MPI_Request request = req_storage.pop(args.src, args.dst, args.tag);
if (request == MPI_REQUEST_NULL) {
/* Assume that the trace is well formed, meaning the comm might have been caught by an MPI_test. Then just
// MPI_REQUEST_NULL by Request::wait!
bool is_wait_for_receive = (request->flags() & MPI_REQ_RECV);
- TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::WaitTIData(args.src, args.dst, args.tag));
+ TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::WaitTIData("wait", args.src, args.dst, args.tag));
MPI_Status status;
Request::wait(&request, &status);
-
+ if(request!=MPI_REQUEST_NULL)
+ Request::unref(&request);
TRACE_smpi_comm_out(get_pid());
if (is_wait_for_receive)
TRACE_smpi_recv(MPI_COMM_WORLD->group()->actor(args.src), MPI_COMM_WORLD->group()->actor(args.dst), args.tag);
void TestAction::kernel(simgrid::xbt::ReplayAction&)
{
const WaitTestParser& args = get_args();
- MPI_Request request = req_storage.find(args.src, args.dst, args.tag);
- req_storage.remove(request);
+ MPI_Request request = req_storage.pop(args.src, args.dst, args.tag);
// if request is null here, this may mean that a previous test has succeeded
// Different times in traced application and replayed version may lead to this
// In this case, ignore the extra calls.
if (request != MPI_REQUEST_NULL) {
- TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::NoOpTIData("test"));
+ TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::WaitTIData("test", args.src, args.dst, args.tag));
MPI_Status status;
int flag = 0;
void WaitAllAction::kernel(simgrid::xbt::ReplayAction&)
{
- const size_t count_requests = req_storage.size();
-
- if (count_requests > 0) {
- TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::CpuTIData("waitall", count_requests));
+ if (req_storage.size() > 0) {
std::vector<std::pair</*sender*/ aid_t, /*recv*/ aid_t>> sender_receiver;
std::vector<MPI_Request> reqs;
req_storage.get_requests(reqs);
+ unsigned long count_requests = reqs.size();
+ TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::CpuTIData("waitall", count_requests));
for (auto const& req : reqs) {
if (req && (req->flags() & MPI_REQ_RECV)) {
sender_receiver.emplace_back(req->src(), req->dst());
Request::waitall(count_requests, &(reqs.data())[0], MPI_STATUSES_IGNORE);
req_storage.get_store().clear();
- for (auto const& pair : sender_receiver) {
- TRACE_smpi_recv(pair.first, pair.second, 0);
+ for (MPI_Request& req : reqs)
+ if (req != MPI_REQUEST_NULL)
+ Request::unref(&req);
+
+ for (auto const& [src, dst] : sender_receiver) {
+ TRACE_smpi_recv(src, dst, 0);
}
TRACE_smpi_comm_out(get_pid());
}
TRACE_smpi_comm_in(
get_pid(), "action_reducescatter",
new simgrid::instr::VarCollTIData(get_name(), -1, -1, nullptr, -1, args.recvcounts,
- std::to_string(args.comp_size),
+ /* ugly as we use datatype field to pass computation as string */
+ /* and because of the trick to avoid getting 0.000000 when 0 is given */
+ args.comp_size == 0 ? "0" : std::to_string(args.comp_size),
Datatype::encode(args.datatype1)));
colls::reduce_scatter(send_buffer(args.recv_size_sum * args.datatype1->size()),
TRACE_smpi_comm_out(get_pid());
}
-} // Replay Namespace
-}} // namespace simgrid::smpi
+} // namespace simgrid::smpi::replay
static std::unordered_map<aid_t, simgrid::smpi::replay::RequestStorage> storage;
/** @brief Only initialize the replay, don't do it for real */
std::vector<MPI_Request> requests(count_requests);
unsigned int i=0;
- for (auto const& pair : storage[simgrid::s4u::this_actor::get_pid()].get_store()) {
- requests[i] = pair.second;
+ for (auto const& [_, reqs] : storage[simgrid::s4u::this_actor::get_pid()].get_store()) {
+ for (auto& req : reqs) {
+ requests[i] = req; // FIXME: overwritten at each iteration?
+ }
i++;
}
simgrid::smpi::Request::waitall(count_requests, requests.data(), MPI_STATUSES_IGNORE);