X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/e1a5484101efbff9c8d982a0c4650f7953ee7e2f..043ad6e05e8fbbff04d21b27b5b369fdbc005e01:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index 97c73047c4..116b5ad907 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -12,13 +12,14 @@ #include #include +#include +#include #include #include +#include #include #include -#include - 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 @@ -61,9 +62,6 @@ public: }; } -using req_key_t = std::tuple; -using req_storage_t = std::unordered_map>>; - void log_timed_action(const simgrid::xbt::ReplayAction& action, double clock) { if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){ @@ -72,10 +70,29 @@ void log_timed_action(const simgrid::xbt::ReplayAction& action, double clock) } } -/* Helper function */ +/* Helper functions */ static double parse_double(const std::string& string) { - return xbt_str_parse_double(string.c_str(), "%s is not a double"); + return xbt_str_parse_double(string.c_str(), "not a double"); +} + +template 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(std::numeric_limits::min()) <= val && + val <= static_cast(std::numeric_limits::max()), + "out of range: %g", val); + return static_cast(val); +} + +static int parse_root(const simgrid::xbt::ReplayAction& action, unsigned i) +{ + return i < action.size() ? std::stoi(action[i]) : 0; +} + +static MPI_Datatype parse_datatype(const simgrid::xbt::ReplayAction& action, unsigned i) +{ + return i < action.size() ? simgrid::smpi::Datatype::decode(action[i]) : simgrid::smpi::replay::MPI_DEFAULT_TYPE; } namespace simgrid { @@ -86,11 +103,14 @@ MPI_Datatype MPI_DEFAULT_TYPE; class RequestStorage { private: - req_storage_t store; + using req_key_t = std::tuple; + using req_storage_t = std::unordered_map>>; + + req_storage_t store; public: RequestStorage() = default; - int size() const { return store.size(); } + size_t size() const { return store.size(); } req_storage_t& get_store() { return store; } @@ -98,7 +118,7 @@ public: { for (auto const& pair : store) { auto& req = pair.second; - auto my_proc_id = simgrid::s4u::this_actor::get_pid(); + 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"); @@ -146,9 +166,8 @@ void SendRecvParser::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_double(action[4]); - if (action.size() > 5) - datatype1 = simgrid::smpi::Datatype::decode(action[5]); + size = parse_integer(action[4]); + datatype1 = parse_datatype(action, 5); } void ComputeParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) @@ -173,42 +192,36 @@ void LocationParser::parse(simgrid::xbt::ReplayAction& action, const std::string void BcastArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { CHECK_ACTION_PARAMS(action, 1, 2) - size = parse_double(action[2]); - root = (action.size() > 3) ? std::stoi(action[3]) : 0; - if (action.size() > 4) - datatype1 = simgrid::smpi::Datatype::decode(action[4]); + size = parse_integer(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) - comm_size = parse_double(action[2]); + comm_size = parse_integer(action[2]); comp_size = parse_double(action[3]); - root = (action.size() > 4) ? std::stoi(action[4]) : 0; - if (action.size() > 5) - datatype1 = simgrid::smpi::Datatype::decode(action[5]); + 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) - comm_size = parse_double(action[2]); + comm_size = parse_integer(action[2]); comp_size = parse_double(action[3]); - if (action.size() > 4) - datatype1 = simgrid::smpi::Datatype::decode(action[4]); + datatype1 = parse_datatype(action, 4); } void AllToAllArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { CHECK_ACTION_PARAMS(action, 2, 1) comm_size = MPI_COMM_WORLD->size(); - send_size = parse_double(action[2]); - recv_size = parse_double(action[3]); - - if (action.size() > 4) - datatype1 = simgrid::smpi::Datatype::decode(action[4]); - if (action.size() > 5) - datatype2 = simgrid::smpi::Datatype::decode(action[5]); + send_size = parse_integer(action[2]); + recv_size = parse_integer(action[3]); + datatype1 = parse_datatype(action, 4); + datatype2 = parse_datatype(action, 5); } void GatherArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string& name) @@ -224,20 +237,17 @@ void GatherArgParser::parse(simgrid::xbt::ReplayAction& action, const std::strin */ 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(action[2]); + recv_size = parse_integer(action[3]); if (name == "gather") { - root = (action.size() > 4) ? std::stoi(action[4]) : 0; - if (action.size() > 5) - datatype1 = simgrid::smpi::Datatype::decode(action[5]); - if (action.size() > 6) - datatype2 = simgrid::smpi::Datatype::decode(action[6]); + root = parse_root(action, 4); + datatype1 = parse_datatype(action, 5); + datatype2 = parse_datatype(action, 6); } else { - if (action.size() > 4) - datatype1 = simgrid::smpi::Datatype::decode(action[4]); - if (action.size() > 5) - datatype2 = simgrid::smpi::Datatype::decode(action[5]); + root = 0; + datatype1 = parse_datatype(action, 4); + datatype2 = parse_datatype(action, 5); } } @@ -254,37 +264,39 @@ void GatherVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::stri */ comm_size = MPI_COMM_WORLD->size(); CHECK_ACTION_PARAMS(action, comm_size + 1, 2) - send_size = parse_double(action[2]); + send_size = parse_integer(action[2]); disps = std::vector(comm_size, 0); recvcounts = std::make_shared>(comm_size); if (name == "gatherv") { - root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; - if (action.size() > 4 + comm_size) - datatype1 = simgrid::smpi::Datatype::decode(action[4 + comm_size]); - if (action.size() > 5 + comm_size) - datatype2 = simgrid::smpi::Datatype::decode(action[5 + comm_size]); + root = parse_root(action, 3 + comm_size); + datatype1 = parse_datatype(action, 4 + comm_size); + datatype2 = parse_datatype(action, 5 + comm_size); } else { - int disp_index = 0; + root = 0; + unsigned disp_index = 0; /* The 3 comes from "0 gather ", which must always be present. * The + comm_size is the recvcounts array, which must also be present */ - if (action.size() > 3 + comm_size + comm_size) { /* datatype + disp are specified */ - int datatype_index = 3 + comm_size; - disp_index = datatype_index + 1; - datatype1 = simgrid::smpi::Datatype::decode(action[datatype_index]); - datatype2 = simgrid::smpi::Datatype::decode(action[datatype_index]); - } else if (action.size() > - 3 + comm_size + 2) { /* disps specified; datatype is not specified; use the default one */ + if (action.size() > 3 + comm_size + comm_size) { + // datatype + disp are specified + datatype1 = parse_datatype(action, 3 + comm_size); + datatype2 = parse_datatype(action, 4 + comm_size); + disp_index = 5 + comm_size; + } else if (action.size() > 3 + comm_size + 2) { + // disps specified; datatype is not specified; use the default one + datatype1 = MPI_DEFAULT_TYPE; + datatype2 = MPI_DEFAULT_TYPE; disp_index = 3 + comm_size; - } else if (action.size() > 3 + comm_size) { /* only datatype, no disp specified */ - int datatype_index = 3 + comm_size; - datatype1 = simgrid::smpi::Datatype::decode(action[datatype_index]); - datatype2 = simgrid::smpi::Datatype::decode(action[datatype_index]); + } else { + // no disp specified, maybe only datatype, + datatype1 = parse_datatype(action, 3 + comm_size); + datatype2 = parse_datatype(action, 4 + comm_size); } if (disp_index != 0) { - for (unsigned int i = 0; i < comm_size; i++) + xbt_assert(disp_index + comm_size <= action.size()); + for (unsigned i = 0; i < comm_size; i++) disps[i] = std::stoi(action[disp_index + i]); } } @@ -308,13 +320,11 @@ void ScatterArgParser::parse(simgrid::xbt::ReplayAction& action, const std::stri */ CHECK_ACTION_PARAMS(action, 2, 3) comm_size = MPI_COMM_WORLD->size(); - send_size = parse_double(action[2]); - recv_size = parse_double(action[3]); - root = (action.size() > 4) ? std::stoi(action[4]) : 0; - if (action.size() > 5) - datatype1 = simgrid::smpi::Datatype::decode(action[5]); - if (action.size() > 6) - datatype2 = simgrid::smpi::Datatype::decode(action[6]); + send_size = parse_integer(action[2]); + recv_size = parse_integer(action[3]); + root = parse_root(action, 4); + datatype1 = parse_datatype(action, 5); + datatype2 = parse_datatype(action, 6); } void ScatterVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) @@ -329,20 +339,18 @@ void ScatterVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::str 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() */ CHECK_ACTION_PARAMS(action, comm_size + 1, 2) - recv_size = parse_double(action[2 + comm_size]); + recv_size = parse_integer(action[2 + comm_size]); disps = std::vector(comm_size, 0); sendcounts = std::make_shared>(comm_size); - if (action.size() > 5 + comm_size) - datatype1 = simgrid::smpi::Datatype::decode(action[4 + comm_size]); - if (action.size() > 5 + comm_size) - datatype2 = simgrid::smpi::Datatype::decode(action[5]); + root = parse_root(action, 3 + comm_size); + datatype1 = parse_datatype(action, 4 + comm_size); + datatype2 = parse_datatype(action, 5 + comm_size); for (unsigned int i = 0; i < comm_size; i++) { (*sendcounts)[i] = std::stoi(action[i + 2]); } send_size_sum = std::accumulate(sendcounts->begin(), sendcounts->end(), 0); - root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; } void ReduceScatterArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) @@ -358,8 +366,7 @@ void ReduceScatterArgParser::parse(simgrid::xbt::ReplayAction& action, const std CHECK_ACTION_PARAMS(action, comm_size + 1, 1) comp_size = parse_double(action[2 + comm_size]); recvcounts = std::make_shared>(comm_size); - if (action.size() > 3 + comm_size) - datatype1 = simgrid::smpi::Datatype::decode(action[3 + 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])); @@ -384,13 +391,11 @@ void AllToAllVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::st senddisps = std::vector(comm_size, 0); recvdisps = std::vector(comm_size, 0); - if (action.size() > 5 + 2 * comm_size) - datatype1 = simgrid::smpi::Datatype::decode(action[4 + 2 * comm_size]); - if (action.size() > 5 + 2 * comm_size) - datatype2 = simgrid::smpi::Datatype::decode(action[5 + 2 * comm_size]); + 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(action[2]); + recv_buf_size = parse_integer(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]); @@ -413,7 +418,7 @@ void WaitAction::kernel(simgrid::xbt::ReplayAction& action) 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! @@ -432,7 +437,7 @@ void WaitAction::kernel(simgrid::xbt::ReplayAction& action) void SendAction::kernel(simgrid::xbt::ReplayAction&) { const SendRecvParser& args = get_args(); - int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner); + aid_t dst_traced = MPI_COMM_WORLD->group()->actor(args.partner); TRACE_smpi_comm_in( get_pid(), __func__, @@ -461,8 +466,8 @@ void RecvAction::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 == 0.0) { Request::probe(args.partner, args.tag, MPI_COMM_WORLD, &status); arg_size = status.count; } @@ -480,7 +485,7 @@ void RecvAction::kernel(simgrid::xbt::ReplayAction&) 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); + aid_t src_traced = MPI_COMM_WORLD->group()->actor(status.MPI_SOURCE); TRACE_smpi_recv(src_traced, get_pid(), args.tag); } } @@ -497,10 +502,10 @@ void SleepAction::kernel(simgrid::xbt::ReplayAction&) { 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&) @@ -553,11 +558,11 @@ void CommunicatorAction::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> sender_receiver; + std::vector> sender_receiver; std::vector reqs; req_storage.get_requests(reqs); for (auto const& req : reqs) { @@ -587,7 +592,7 @@ void BcastAction::kernel(simgrid::xbt::ReplayAction&) 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, - -1, Datatype::encode(args.datatype1), "")); + 0, Datatype::encode(args.datatype1), "")); colls::bcast(send_buffer(args.size * args.datatype1->size()), args.size, args.datatype1, args.root, MPI_COMM_WORLD); @@ -599,7 +604,7 @@ void ReduceAction::kernel(simgrid::xbt::ReplayAction&) 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, - args.comm_size, -1, Datatype::encode(args.datatype1), "")); + 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, @@ -613,7 +618,7 @@ void AllReduceAction::kernel(simgrid::xbt::ReplayAction&) { 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()), @@ -632,8 +637,8 @@ void AllToAllAction::kernel(simgrid::xbt::ReplayAction&) Datatype::encode(args.datatype1), Datatype::encode(args.datatype2))); - colls::alltoall(send_buffer(args.send_size * args.comm_size * args.datatype1->size()), args.send_size, args.datatype1, - recv_buffer(args.recv_size * args.comm_size * args.datatype2->size()), args.recv_size, args.datatype2, + colls::alltoall(send_buffer(args.datatype1->size() * args.send_size * args.comm_size), args.send_size, args.datatype1, + recv_buffer(args.datatype2->size() * args.recv_size * args.comm_size), args.recv_size, args.datatype2, MPI_COMM_WORLD); TRACE_smpi_comm_out(get_pid()); @@ -650,7 +655,7 @@ void GatherAction::kernel(simgrid::xbt::ReplayAction&) if (get_name() == "gather") { int rank = MPI_COMM_WORLD->rank(); colls::gather(send_buffer(args.send_size * args.datatype1->size()), args.send_size, args.datatype1, - (rank == args.root) ? recv_buffer(args.recv_size * args.comm_size * args.datatype2->size()) : nullptr, + (rank == args.root) ? recv_buffer(args.datatype2->size() * args.recv_size * args.comm_size) : nullptr, args.recv_size, args.datatype2, args.root, MPI_COMM_WORLD); } else colls::allgather(send_buffer(args.send_size * args.datatype1->size()), args.send_size, args.datatype1, @@ -666,7 +671,7 @@ void GatherVAction::kernel(simgrid::xbt::ReplayAction&) 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") { @@ -703,7 +708,7 @@ void ScatterVAction::kernel(simgrid::xbt::ReplayAction&) 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))); @@ -720,7 +725,7 @@ void ReduceScatterAction::kernel(simgrid::xbt::ReplayAction&) 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))); @@ -737,7 +742,7 @@ void AllToAllVAction::kernel(simgrid::xbt::ReplayAction&) 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(), @@ -762,9 +767,7 @@ void smpi_replay_init(const char* instance_id, int rank, double start_delay_flop 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); }); @@ -816,7 +819,7 @@ void smpi_replay_main(int rank, const char* 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) { std::vector requests(count_requests); unsigned int i=0;