-/* Copyright (c) 2009-2020. The SimGrid Team. All rights reserved. */
+/* Copyright (c) 2009-2021. 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_group.hpp"
#include "smpi_request.hpp"
#include "xbt/replay.hpp"
-#include <simgrid/smpi/replay.hpp>
+#include <simgrid/smpi/smpi_replay.hpp>
#include <src/smpi/include/private.hpp>
+#include <cmath>
+#include <limits>
#include <memory>
#include <numeric>
+#include <tuple>
#include <unordered_map>
#include <vector>
-#include <tuple>
-
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
};
}
-typedef std::tuple</*sender*/ int, /* receiver */ int, /* tag */ int> req_key_t;
-typedef std::unordered_map<req_key_t, MPI_Request, hash_tuple::hash<std::tuple<int,int,int>>> req_storage_t;
-
void log_timed_action(const simgrid::xbt::ReplayAction& action, double clock)
{
if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){
class RequestStorage {
private:
- req_storage_t store;
+ 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>>>;
+
+ req_storage_t store;
public:
- RequestStorage() {}
- int size() const { return store.size(); }
+ RequestStorage() = default;
+ size_t size() const { return store.size(); }
- req_storage_t& get_store()
- {
- return store;
- }
+ req_storage_t& get_store() { return store; }
- void get_requests(std::vector<MPI_Request>& vec) const
- {
- for (auto const& pair : store) {
- auto& req = pair.second;
- auto 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");
- }
+ void get_requests(std::vector<MPI_Request>& vec) const
+ {
+ for (auto const& pair : store) {
+ auto& req = pair.second;
+ 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");
}
}
+ }
MPI_Request find(int src, int dst, int tag)
{
- req_storage_t::iterator it = store.find(req_key_t(src, dst, tag));
+ auto it = store.find(req_key_t(src, dst, tag));
return (it == store.end()) ? MPI_REQUEST_NULL : it->second;
}
/* 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)->get_pid()-1,
- MPI_COMM_WORLD->group()->actor(dst)->get_pid()-1,
- tag), MPI_REQUEST_NULL});
+ store.insert({req_key_t(MPI_COMM_WORLD->group()->actor(src) - 1, MPI_COMM_WORLD->group()->actor(dst) - 1, tag),
+ MPI_REQUEST_NULL});
}
};
void ReduceArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 2, 2)
- comm_size = parse_double(action[2]);
+ double arg2 = trunc(parse_double(action[2]));
+ xbt_assert(0.0 <= arg2 && arg2 <= static_cast<double>(std::numeric_limits<unsigned>::max()));
+ comm_size = static_cast<unsigned>(arg2);
comp_size = parse_double(action[3]);
root = (action.size() > 4) ? std::stoi(action[4]) : 0;
if (action.size() > 5)
void AllReduceArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 2, 1)
- comm_size = parse_double(action[2]);
+ double arg2 = trunc(parse_double(action[2]));
+ xbt_assert(0.0 <= arg2 && arg2 <= static_cast<double>(std::numeric_limits<unsigned>::max()));
+ comm_size = static_cast<unsigned>(arg2);
comp_size = parse_double(action[3]);
if (action.size() > 4)
datatype1 = simgrid::smpi::Datatype::decode(action[4]);
req_storage.remove(request);
if (request == MPI_REQUEST_NULL) {
- /* Assume that the trace is well formed, meaning the comm might have been caught by a MPI_test. Then just
+ /* Assume that the trace is well formed, meaning the comm might have been caught by an MPI_test. Then just
* return.*/
return;
}
- int rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1;
+ aid_t rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1;
// Must be taken before Request::wait() since the request may be set to
// MPI_REQUEST_NULL by Request::wait!
void SendAction::kernel(simgrid::xbt::ReplayAction&)
{
const SendRecvParser& args = get_args();
- int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid();
+ aid_t dst_traced = MPI_COMM_WORLD->group()->actor(args.partner);
TRACE_smpi_comm_in(
get_pid(), __func__,
TRACE_smpi_comm_out(get_pid());
if (is_recv && not TRACE_smpi_view_internals()) {
- int src_traced = MPI_COMM_WORLD->group()->actor(status.MPI_SOURCE)->get_pid();
+ aid_t src_traced = MPI_COMM_WORLD->group()->actor(status.MPI_SOURCE);
TRACE_smpi_recv(src_traced, get_pid(), args.tag);
}
}
{
const SleepParser& args = get_args();
XBT_DEBUG("Sleep for: %lf secs", args.time);
- int rank = simgrid::s4u::this_actor::get_pid();
- TRACE_smpi_sleeping_in(rank, args.time);
+ aid_t pid = simgrid::s4u::this_actor::get_pid();
+ TRACE_smpi_sleeping_in(pid, args.time);
simgrid::s4u::this_actor::sleep_for(args.time/smpi_adjust_comp_speed());
- TRACE_smpi_sleeping_out(rank);
+ TRACE_smpi_sleeping_out(pid);
}
void LocationAction::kernel(simgrid::xbt::ReplayAction&)
void WaitAllAction::kernel(simgrid::xbt::ReplayAction&)
{
- const unsigned int count_requests = req_storage.size();
+ const size_t count_requests = req_storage.size();
if (count_requests > 0) {
TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::Pt2PtTIData("waitall", -1, count_requests, ""));
- std::vector<std::pair</*sender*/int,/*recv*/int>> sender_receiver;
+ std::vector<std::pair</*sender*/ aid_t, /*recv*/ aid_t>> sender_receiver;
std::vector<MPI_Request> reqs;
req_storage.get_requests(reqs);
for (auto const& req : reqs) {
if (req && (req->flags() & MPI_REQ_RECV)) {
- sender_receiver.push_back({req->src(), req->dst()});
+ sender_receiver.emplace_back(req->src(), req->dst());
}
}
Request::waitall(count_requests, &(reqs.data())[0], MPI_STATUSES_IGNORE);
{
const BcastArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), "action_bcast",
- new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->actor(args.root)->get_pid(), -1.0,
- args.size, -1, Datatype::encode(args.datatype1), ""));
+ new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->actor(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()->actor(args.root)->get_pid(),
- args.comp_size, args.comm_size, -1,
- Datatype::encode(args.datatype1), ""));
+ new simgrid::instr::CollTIData("reduce", MPI_COMM_WORLD->group()->actor(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,
{
const AllReduceArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), "action_allreduce",
- new simgrid::instr::CollTIData("allreduce", -1, args.comp_size, args.comm_size, -1,
+ new simgrid::instr::CollTIData("allreduce", -1, args.comp_size, args.comm_size, 0,
Datatype::encode(args.datatype1), ""));
colls::allreduce(send_buffer(args.comm_size * args.datatype1->size()),
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, -1,
+ get_name(), (get_name() == "gatherv") ? args.root : -1, args.send_size, nullptr, 0,
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, -1, args.sendcounts, args.recv_size,
+ new simgrid::instr::VarCollTIData(get_name(), args.root, 0, 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, -1, args.recvcounts,
+ new simgrid::instr::VarCollTIData("reducescatter", -1, 0, nullptr, 0, args.recvcounts,
std::to_string(args.comp_size), /* ugly hack to print comp_size */
Datatype::encode(args.datatype1)));
const AllToAllVArgParser& args = get_args();
TRACE_smpi_comm_in(get_pid(), __func__,
new simgrid::instr::VarCollTIData(
- "alltoallv", -1, args.send_size_sum, args.sendcounts, args.recv_size_sum, args.recvcounts,
+ "alltoallv", 0, 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(),
smpi_process()->mark_as_initialized();
smpi_process()->set_replaying(true);
- int my_proc_id = simgrid::s4u::this_actor::get_pid();
-
- TRACE_smpi_init(my_proc_id, "smpi_replay_run_init");
+ TRACE_smpi_init(simgrid::s4u::this_actor::get_pid(), "smpi_replay_run_init");
xbt_replay_action_register("init", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::InitAction().execute(action); });
xbt_replay_action_register("finalize", [](simgrid::xbt::ReplayAction const&) { /* nothing to do */ });
xbt_replay_action_register("comm_size", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::CommunicatorAction().execute(action); });
}
/** @brief actually run the replay after initialization */
-void smpi_replay_main(int rank, const char* trace_filename)
+void smpi_replay_main(int rank, const char* private_trace_filename)
{
static int active_processes = 0;
active_processes++;
storage[simgrid::s4u::this_actor::get_pid()] = simgrid::smpi::replay::RequestStorage();
std::string rank_string = std::to_string(rank);
- simgrid::xbt::replay_runner(rank_string.c_str(), trace_filename);
+ simgrid::xbt::replay_runner(rank_string.c_str(), private_trace_filename);
/* and now, finalize everything */
/* One active process will stop. Decrease the counter*/
unsigned int count_requests = storage[simgrid::s4u::this_actor::get_pid()].size();
- XBT_DEBUG("There are %ud elements in reqq[*]", count_requests);
+ XBT_DEBUG("There are %u elements in reqq[*]", count_requests);
if (count_requests > 0) {
- auto* requests = new MPI_Request[count_requests];
+ 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;
i++;
}
- simgrid::smpi::Request::waitall(count_requests, requests, MPI_STATUSES_IGNORE);
- delete[] requests;
+ simgrid::smpi::Request::waitall(count_requests, requests.data(), MPI_STATUSES_IGNORE);
}
active_processes--;
}
/** @brief chain a replay initialization and a replay start */
-void smpi_replay_run(const char* instance_id, int rank, double start_delay_flops, const char* trace_filename)
+void smpi_replay_run(const char* instance_id, int rank, double start_delay_flops, const char* private_trace_filename)
{
smpi_replay_init(instance_id, rank, start_delay_flops);
- smpi_replay_main(rank, trace_filename);
+ smpi_replay_main(rank, private_trace_filename);
}