-/* Copyright (c) 2009-2021. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2009-2023. 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 "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.
}
// Recursive template code derived from Matthieu M.
-template <class Tuple, size_t Index = std::tuple_size<Tuple>::value - 1> class HashValueImpl {
+template <class Tuple, size_t Index = std::tuple_size_v<Tuple> - 1> class HashValueImpl {
public:
static void apply(size_t& seed, Tuple const& tuple)
{
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 (const 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&)
tag = std::stoi(action[4]);
}
-void SendRecvParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
+void SendOrRecvParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 3, 1)
partner = std::stoi(action[2]);
tag = std::stoi(action[3]);
- size = parse_integer<size_t>(action[4]);
+ size = parse_integer<ssize_t>(action[4]);
datatype1 = parse_datatype(action, 5);
}
void LocationParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 2, 0)
- filename = std::string(action[2]);
+ filename = action[2];
line = std::stoi(action[3]);
}
+void SendRecvParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
+{
+ CHECK_ACTION_PARAMS(action, 6, 0)
+ sendcount = parse_integer<int>(action[2]);
+ dst = std::stoi(action[3]);
+ recvcount = parse_integer<int>(action[4]);
+ src = std::stoi(action[5]);
+ datatype1 = parse_datatype(action, 6);
+ datatype2 = parse_datatype(action, 7);
+}
+
void BcastArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 1, 2)
4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode()
5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode()
*/
+ comm_size = MPI_COMM_WORLD->size();
CHECK_ACTION_PARAMS(action, 2, 3)
comm_size = MPI_COMM_WORLD->size();
send_size = parse_integer<int>(action[2]);
4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode()
5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode()
*/
+ comm_size = MPI_COMM_WORLD->size();
CHECK_ACTION_PARAMS(action, comm_size + 1, 2)
recv_size = parse_integer<int>(action[2 + comm_size]);
disps = std::vector<int>(comm_size, 0);
datatype1 = parse_datatype(action, 3 + comm_size);
for (unsigned int i = 0; i < comm_size; i++) {
- recvcounts->push_back(std::stoi(action[i + 2]));
+ (*recvcounts)[i]= std::stoi(action[i + 2]);
}
recv_size_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0);
}
+void ScanArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
+{
+ CHECK_ACTION_PARAMS(action, 2, 1)
+ size = parse_integer<size_t>(action[2]);
+ comp_size = parse_double(action[3]);
+ datatype1 = parse_datatype(action, 4);
+}
+
void AllToAllVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
/* The structure of the alltoallv action for the rank 0 (total 4 processes) is the following:
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 SendAction::kernel(simgrid::xbt::ReplayAction&)
{
- const SendRecvParser& args = get_args();
+ const SendOrRecvParser& args = get_args();
aid_t dst_traced = MPI_COMM_WORLD->group()->actor(args.partner);
TRACE_smpi_comm_in(
void RecvAction::kernel(simgrid::xbt::ReplayAction&)
{
- const SendRecvParser& args = get_args();
+ const SendOrRecvParser& args = get_args();
TRACE_smpi_comm_in(
get_pid(), __func__,
new simgrid::instr::Pt2PtTIData(get_name(), args.partner, args.size, args.tag, Datatype::encode(args.datatype1)));
MPI_Status status;
// unknown size from the receiver point of view
- size_t arg_size = args.size;
- if (arg_size == 0) {
+ ssize_t arg_size = args.size;
+ if (arg_size < 0) {
Request::probe(args.partner, args.tag, MPI_COMM_WORLD, &status);
arg_size = status.count;
}
}
}
+void SendRecvAction::kernel(simgrid::xbt::ReplayAction&)
+{
+ XBT_DEBUG("Enters SendRecv");
+ const SendRecvParser& args = get_args();
+ aid_t my_proc_id = simgrid::s4u::this_actor::get_pid();
+ aid_t src_traced = MPI_COMM_WORLD->group()->actor(args.src);
+ aid_t dst_traced = MPI_COMM_WORLD->group()->actor(args.dst);
+
+ MPI_Status status;
+ int sendtag=0;
+ int recvtag=0;
+
+ // FIXME: Hack the way to trace this one
+ auto dst_hack = std::make_shared<std::vector<int>>();
+ auto src_hack = std::make_shared<std::vector<int>>();
+ dst_hack->push_back(dst_traced);
+ src_hack->push_back(src_traced);
+ TRACE_smpi_comm_in(my_proc_id, __func__,
+ new simgrid::instr::VarCollTIData(
+ "sendRecv", -1, args.sendcount,
+ dst_hack, args.recvcount, src_hack,
+ simgrid::smpi::Datatype::encode(args.datatype1), simgrid::smpi::Datatype::encode(args.datatype2)));
+
+ TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, sendtag, args.sendcount * args.datatype1->size());
+
+ simgrid::smpi::Request::sendrecv(nullptr, args.sendcount, args.datatype1, args.dst, sendtag, nullptr, args.recvcount, args.datatype2, args.src,
+ recvtag, MPI_COMM_WORLD, &status);
+
+ TRACE_smpi_recv(src_traced, my_proc_id, recvtag);
+ TRACE_smpi_comm_out(my_proc_id);
+ XBT_DEBUG("Exits SendRecv");
+}
+
void ComputeAction::kernel(simgrid::xbt::ReplayAction&)
{
const ComputeParser& args = get_args();
void LocationAction::kernel(simgrid::xbt::ReplayAction&)
{
const LocationParser& args = get_args();
- smpi_trace_set_call_location(args.filename.c_str(), args.line);
+ smpi_trace_set_call_location(args.filename.c_str(), args.line, "replay_action");
}
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());
}
{
const BcastArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), "action_bcast",
- new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->rank(args.root), -1.0, args.size,
+ new simgrid::instr::CollTIData("bcast", args.root, -1.0, args.size,
0, Datatype::encode(args.datatype1), ""));
colls::bcast(send_buffer(args.size * args.datatype1->size()), args.size, args.datatype1, args.root, MPI_COMM_WORLD);
{
const ReduceArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), "action_reduce",
- new simgrid::instr::CollTIData("reduce", MPI_COMM_WORLD->group()->rank(args.root), args.comp_size,
+ new simgrid::instr::CollTIData("reduce", args.root, args.comp_size,
args.comm_size, 0, Datatype::encode(args.datatype1), ""));
colls::reduce(send_buffer(args.comm_size * args.datatype1->size()),
recv_buffer(args.comm_size * args.datatype1->size()), args.comm_size, args.datatype1, MPI_OP_NULL,
args.root, MPI_COMM_WORLD);
- if(args.comp_size != 0.0)
- private_execute_flops(args.comp_size);
+ if (args.comp_size != 0.0)
+ simgrid::s4u::this_actor::exec_init(args.comp_size)
+ ->set_name("computation")
+ ->start()
+ ->wait();
TRACE_smpi_comm_out(get_pid());
}
colls::allreduce(send_buffer(args.comm_size * args.datatype1->size()),
recv_buffer(args.comm_size * args.datatype1->size()), args.comm_size, args.datatype1, MPI_OP_NULL,
MPI_COMM_WORLD);
- if(args.comp_size != 0.0)
- private_execute_flops(args.comp_size);
+ if (args.comp_size != 0.0)
+ simgrid::s4u::this_actor::exec_init(args.comp_size)
+ ->set_name("computation")
+ ->start()
+ ->wait();
TRACE_smpi_comm_out(get_pid());
}
const GatherVArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), get_name().c_str(),
new simgrid::instr::VarCollTIData(
- get_name(), (get_name() == "gatherv") ? args.root : -1, args.send_size, nullptr, 0,
+ get_name(), (get_name() == "gatherv") ? args.root : -1, args.send_size, nullptr, -1,
args.recvcounts, Datatype::encode(args.datatype1), Datatype::encode(args.datatype2)));
if (get_name() == "gatherv") {
int rank = MPI_COMM_WORLD->rank();
const ScatterVArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), "action_scatterv",
- new simgrid::instr::VarCollTIData(get_name(), args.root, 0, args.sendcounts, args.recv_size,
+ new simgrid::instr::VarCollTIData(get_name(), args.root, -1, args.sendcounts, args.recv_size,
nullptr, Datatype::encode(args.datatype1),
Datatype::encode(args.datatype2)));
const ReduceScatterArgParser& args = get_args();
TRACE_smpi_comm_in(
get_pid(), "action_reducescatter",
- new simgrid::instr::VarCollTIData("reducescatter", -1, 0, nullptr, 0, args.recvcounts,
- std::to_string(args.comp_size), /* ugly hack to print comp_size */
+ new simgrid::instr::VarCollTIData(get_name(), -1, -1, nullptr, -1, args.recvcounts,
+ /* 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()),
recv_buffer(args.recv_size_sum * args.datatype1->size()), args.recvcounts->data(),
args.datatype1, MPI_OP_NULL, MPI_COMM_WORLD);
+ if (args.comp_size != 0.0)
+ simgrid::s4u::this_actor::exec_init(args.comp_size)
+ ->set_name("computation")
+ ->start()
+ ->wait();
+ TRACE_smpi_comm_out(get_pid());
+}
- private_execute_flops(args.comp_size);
+void ScanAction::kernel(simgrid::xbt::ReplayAction&)
+{
+ const ScanArgParser& args = get_args();
+ TRACE_smpi_comm_in(get_pid(), "action_scan",
+ new simgrid::instr::CollTIData(get_name(), -1, args.comp_size,
+ args.size, 0, Datatype::encode(args.datatype1), ""));
+ if (get_name() == "scan")
+ colls::scan(send_buffer(args.size * args.datatype1->size()),
+ recv_buffer(args.size * args.datatype1->size()), args.size,
+ args.datatype1, MPI_OP_NULL, MPI_COMM_WORLD);
+ else
+ colls::exscan(send_buffer(args.size * args.datatype1->size()),
+ recv_buffer(args.size * args.datatype1->size()), args.size,
+ args.datatype1, MPI_OP_NULL, MPI_COMM_WORLD);
+
+ if (args.comp_size != 0.0)
+ simgrid::s4u::this_actor::exec_init(args.comp_size)
+ ->set_name("computation")
+ ->start()
+ ->wait();
TRACE_smpi_comm_out(get_pid());
}
const AllToAllVArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), __func__,
new simgrid::instr::VarCollTIData(
- "alltoallv", 0, args.send_size_sum, args.sendcounts, args.recv_size_sum, args.recvcounts,
+ "alltoallv", -1, args.send_size_sum, args.sendcounts, args.recv_size_sum, args.recvcounts,
Datatype::encode(args.datatype1), Datatype::encode(args.datatype2)));
colls::alltoallv(send_buffer(args.send_buf_size * args.datatype1->size()), args.sendcounts->data(),
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 */
xbt_replay_action_register("recv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::RecvAction("recv", storage[simgrid::s4u::this_actor::get_pid()]).execute(action); });
xbt_replay_action_register("irecv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::RecvAction("irecv", storage[simgrid::s4u::this_actor::get_pid()]).execute(action); });
xbt_replay_action_register("test", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::TestAction(storage[simgrid::s4u::this_actor::get_pid()]).execute(action); });
+ xbt_replay_action_register("sendRecv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SendRecvAction().execute(action); });
xbt_replay_action_register("wait", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::WaitAction(storage[simgrid::s4u::this_actor::get_pid()]).execute(action); });
xbt_replay_action_register("waitall", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::WaitAllAction(storage[simgrid::s4u::this_actor::get_pid()]).execute(action); });
xbt_replay_action_register("barrier", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::BarrierAction().execute(action); });
xbt_replay_action_register("allgather", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::GatherAction("allgather").execute(action); });
xbt_replay_action_register("allgatherv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::GatherVAction("allgatherv").execute(action); });
xbt_replay_action_register("reducescatter", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ReduceScatterAction().execute(action); });
+ xbt_replay_action_register("scan", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ScanAction("scan").execute(action); });
+ xbt_replay_action_register("exscan", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ScanAction("exscan").execute(action); });
xbt_replay_action_register("compute", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ComputeAction().execute(action); });
xbt_replay_action_register("sleep", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SleepAction().execute(action); });
xbt_replay_action_register("location", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::LocationAction().execute(action); });
// Wait for the other actors to initialize also
simgrid::s4u::this_actor::yield();
}
+ if(_smpi_init_sleep > 0)
+ simgrid::s4u::this_actor::sleep_for(_smpi_init_sleep);
}
/** @brief actually run the replay after initialization */
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 (const auto& req : reqs) {
+ requests[i] = req; // FIXME: overwritten at each iteration?
+ }
i++;
}
simgrid::smpi::Request::waitall(count_requests, requests.data(), MPI_STATUSES_IGNORE);
}
- if(simgrid::config::get_value<bool>("smpi/finalization-barrier"))
+ if (simgrid::config::get_value<bool>("smpi/barrier-finalization"))
simgrid::smpi::colls::barrier(MPI_COMM_WORLD);
active_processes--;
