X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/42f80f7c0ce00e583f9c787127492afecd17419e..HEAD:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index 41cca338ce..701323fb28 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -1,26 +1,29 @@ -/* 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 -#include +#include "xbt/str.h" +#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 // 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. @@ -36,7 +39,7 @@ template inline void hash_combine(std::size_t& seed, T const& v) } // Recursive template code derived from Matthieu M. -template ::value - 1> class HashValueImpl { +template - 1> class HashValueImpl { public: static void apply(size_t& seed, Tuple const& tuple) { @@ -61,9 +64,6 @@ 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)){ @@ -72,65 +72,86 @@ 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); } -namespace simgrid { -namespace smpi { +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 replay { +namespace simgrid::smpi::replay { 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, hash_tuple::hash>>; + + 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; } void get_requests(std::vector& 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&) @@ -141,14 +162,13 @@ 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_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&) @@ -166,49 +186,54 @@ void SleepParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) 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(action[2]); + dst = std::stoi(action[3]); + recvcount = parse_integer(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) - 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 +249,17 @@ 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 +276,39 @@ 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]); } } @@ -306,15 +330,14 @@ void ScatterArgParser::parse(simgrid::xbt::ReplayAction& action, const std::stri 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]); - 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&) @@ -328,21 +351,20 @@ void ScatterVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::str 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(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&) @@ -358,15 +380,22 @@ 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])); + (*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(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: @@ -384,13 +413,11 @@ 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]); @@ -404,8 +431,7 @@ void WaitAction::kernel(simgrid::xbt::ReplayAction& action) 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 @@ -413,25 +439,24 @@ void WaitAction::kernel(simgrid::xbt::ReplayAction& action) 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("wait", args.src, args.dst, args.tag)); MPI_Status status; Request::wait(&request, &status); - - TRACE_smpi_comm_out(rank); + if(request!=MPI_REQUEST_NULL) + Request::unref(&request); + 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&) { - 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( @@ -454,15 +479,15 @@ void SendAction::kernel(simgrid::xbt::ReplayAction&) 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 - double arg_size = args.size; - if (arg_size <= 0.0) { + ssize_t arg_size = args.size; + if (arg_size < 0) { Request::probe(args.partner, args.tag, MPI_COMM_WORLD, &status); arg_size = status.count; } @@ -485,6 +510,39 @@ void RecvAction::kernel(simgrid::xbt::ReplayAction&) } } +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>(); + auto src_hack = std::make_shared>(); + 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(); @@ -506,19 +564,18 @@ void SleepAction::kernel(simgrid::xbt::ReplayAction&) 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; @@ -553,13 +610,12 @@ void CommunicatorAction::kernel(simgrid::xbt::ReplayAction&) 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::Pt2PtTIData("waitall", -1, count_requests, "")); + if (req_storage.size() > 0) { std::vector> sender_receiver; std::vector 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()); @@ -568,8 +624,12 @@ void WaitAllAction::kernel(simgrid::xbt::ReplayAction&) 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()); } @@ -586,8 +646,8 @@ 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), "")); + 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); @@ -598,13 +658,17 @@ 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), "")); + 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()); } @@ -613,13 +677,17 @@ 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()), 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()); } @@ -632,8 +700,8 @@ 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 +718,7 @@ 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 +734,7 @@ 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, 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") { @@ -703,7 +771,7 @@ 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, 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))); @@ -720,15 +788,43 @@ 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, 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()); } @@ -737,7 +833,7 @@ void AllToAllVAction::kernel(simgrid::xbt::ReplayAction&) 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(), @@ -746,8 +842,7 @@ void AllToAllVAction::kernel(simgrid::xbt::ReplayAction&) TRACE_smpi_comm_out(get_pid()); } -} // Replay Namespace -}} // namespace simgrid::smpi +} // namespace simgrid::smpi::replay static std::unordered_map storage; /** @brief Only initialize the replay, don't do it for real */ @@ -773,6 +868,7 @@ void smpi_replay_init(const char* instance_id, int rank, double start_delay_flop 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); }); @@ -788,6 +884,8 @@ void smpi_replay_init(const char* instance_id, int rank, double start_delay_flop 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); }); @@ -800,6 +898,8 @@ void smpi_replay_init(const char* instance_id, int rank, double start_delay_flop // 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 */ @@ -814,17 +914,23 @@ 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; - 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("smpi/barrier-finalization")) + simgrid::smpi::colls::barrier(MPI_COMM_WORLD); + active_processes--; if(active_processes==0){