/* 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 <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 xbt_str_parse_double(string.c_str(), "not a double");
}
+template <typename T> static T parse_integer(const std::string& string)
+{
+ double val = trunc(xbt_str_parse_double(string.c_str(), "not a double"));
+ xbt_assert(static_cast<double>(std::numeric_limits<T>::min()) <= val &&
+ val <= static_cast<double>(std::numeric_limits<T>::max()),
+ "out of range: %g", val);
+ return static_cast<T>(val);
+}
+
static int parse_root(const simgrid::xbt::ReplayAction& action, unsigned i)
{
return i < action.size() ? std::stoi(action[i]) : 0;
CHECK_ACTION_PARAMS(action, 3, 1)
partner = std::stoi(action[2]);
tag = std::stoi(action[3]);
- size = parse_double(action[4]);
+ size = parse_integer<size_t>(action[4]);
datatype1 = parse_datatype(action, 5);
}
void BcastArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 1, 2)
- size = parse_double(action[2]);
+ size = parse_integer<size_t>(action[2]);
root = parse_root(action, 3);
datatype1 = parse_datatype(action, 4);
}
void ReduceArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 2, 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);
+ comm_size = parse_integer<unsigned>(action[2]);
comp_size = parse_double(action[3]);
root = parse_root(action, 4);
datatype1 = parse_datatype(action, 5);
void AllReduceArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&)
{
CHECK_ACTION_PARAMS(action, 2, 1)
- 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);
+ comm_size = parse_integer<unsigned>(action[2]);
comp_size = parse_double(action[3]);
datatype1 = parse_datatype(action, 4);
}
{
CHECK_ACTION_PARAMS(action, 2, 1)
comm_size = MPI_COMM_WORLD->size();
- send_size = parse_double(action[2]);
- recv_size = parse_double(action[3]);
+ send_size = parse_integer<int>(action[2]);
+ recv_size = parse_integer<int>(action[3]);
datatype1 = parse_datatype(action, 4);
datatype2 = parse_datatype(action, 5);
}
*/
CHECK_ACTION_PARAMS(action, 2, 3)
comm_size = MPI_COMM_WORLD->size();
- send_size = parse_double(action[2]);
- recv_size = parse_double(action[3]);
+ send_size = parse_integer<int>(action[2]);
+ recv_size = parse_integer<int>(action[3]);
if (name == "gather") {
root = parse_root(action, 4);
*/
comm_size = MPI_COMM_WORLD->size();
CHECK_ACTION_PARAMS(action, comm_size + 1, 2)
- send_size = parse_double(action[2]);
+ send_size = parse_integer<int>(action[2]);
disps = std::vector<int>(comm_size, 0);
recvcounts = std::make_shared<std::vector<int>>(comm_size);
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_double(action[2]);
- recv_size = parse_double(action[3]);
+ send_size = parse_integer<int>(action[2]);
+ recv_size = parse_integer<int>(action[3]);
root = parse_root(action, 4);
datatype1 = parse_datatype(action, 5);
datatype2 = parse_datatype(action, 6);
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_double(action[2 + comm_size]);
+ recv_size = parse_integer<int>(action[2 + comm_size]);
disps = std::vector<int>(comm_size, 0);
sendcounts = std::make_shared<std::vector<int>>(comm_size);
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:
datatype1 = parse_datatype(action, 4 + 2 * comm_size);
datatype2 = parse_datatype(action, 5 + 2 * comm_size);
- send_buf_size = parse_double(action[2]);
- recv_buf_size = parse_double(action[3 + comm_size]);
+ send_buf_size = parse_integer<int>(action[2]);
+ recv_buf_size = parse_integer<int>(action[3 + comm_size]);
for (unsigned int i = 0; i < comm_size; i++) {
(*sendcounts)[i] = std::stoi(action[3 + i]);
(*recvcounts)[i] = std::stoi(action[4 + comm_size + i]);
return;
}
- 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!
bool is_wait_for_receive = (request->flags() & MPI_REQ_RECV);
- // TODO: Here we take the rank while we normally take the process id (look for get_pid())
- TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::WaitTIData(args.src, args.dst, args.tag));
+
+ TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::WaitTIData(args.src, args.dst, args.tag));
MPI_Status status;
Request::wait(&request, &status);
- TRACE_smpi_comm_out(rank);
+ TRACE_smpi_comm_out(get_pid());
if (is_wait_for_receive)
- TRACE_smpi_recv(args.src, args.dst, args.tag);
+ 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&)
MPI_Status status;
// unknown size from the receiver point of view
- double arg_size = args.size;
- if (arg_size <= 0.0) {
+ size_t arg_size = args.size;
+ if (arg_size == 0) {
Request::probe(args.partner, args.tag, MPI_COMM_WORLD, &status);
arg_size = status.count;
}
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, ""));
+ TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::CpuTIData("waitall", count_requests));
std::vector<std::pair</*sender*/ aid_t, /*recv*/ aid_t>> sender_receiver;
std::vector<MPI_Request> reqs;
req_storage.get_requests(reqs);
{
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), -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()->actor(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);
- 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);
- 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(),
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 */
}
simgrid::smpi::Request::waitall(count_requests, requests.data(), MPI_STATUSES_IGNORE);
}
+
+ if(simgrid::config::get_value<bool>("smpi/finalization-barrier"))
+ simgrid::smpi::colls::barrier(MPI_COMM_WORLD);
+
active_processes--;
if(active_processes==0){