X-Git-Url: http://bilbo.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/05235cab14333e478e8719a57ba7b89fc0d30139..f99dc20a7a829278a448ddf412abe4be587807f5:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index 8067fd4cd3..51e0992e62 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -1,896 +1,941 @@ -/* Copyright (c) 2009-2017. 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 "private.hpp" +#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_process.hpp" #include "smpi_request.hpp" +#include "src/smpi/include/private.hpp" #include "xbt/replay.hpp" +#include "xbt/str.h" -#include +#include +#include +#include #include #include +#include #include #include -using simgrid::s4u::Actor; +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. +namespace hash_tuple { +template class hash { +public: + size_t operator()(TT const& tt) const { return std::hash()(tt); } +}; -XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI"); +template inline void hash_combine(std::size_t& seed, T const& v) +{ + seed ^= hash_tuple::hash()(v) + 0x9e3779b9 + (seed << 6) + (seed >> 2); +} -static int communicator_size = 0; -static int active_processes = 0; -static std::unordered_map*> reqq; +// Recursive template code derived from Matthieu M. +template - 1> class HashValueImpl { +public: + static void apply(size_t& seed, Tuple const& tuple) + { + HashValueImpl::apply(seed, tuple); + hash_combine(seed, std::get(tuple)); + } +}; -static MPI_Datatype MPI_DEFAULT_TYPE; +template class HashValueImpl { +public: + static void apply(size_t& seed, Tuple const& tuple) { hash_combine(seed, std::get<0>(tuple)); } +}; -#define CHECK_ACTION_PARAMS(action, mandatory, optional) \ - { \ - if (action.size() < static_cast(mandatory + 2)) \ - THROWF(arg_error, 0, "%s replay failed.\n" \ - "%lu items were given on the line. First two should be process_id and action. " \ - "This action needs after them %lu mandatory arguments, and accepts %lu optional ones. \n" \ - "Please contact the Simgrid team if support is needed", \ - __FUNCTION__, action.size(), static_cast(mandatory), \ - static_cast(optional)); \ +template class hash> { +public: + size_t operator()(std::tuple const& tt) const + { + size_t seed = 0; + HashValueImpl>::apply(seed, tt); + return seed; } +}; +} -static void log_timed_action (simgrid::xbt::ReplayAction& action, double clock){ +void log_timed_action(const simgrid::xbt::ReplayAction& action, double clock) +{ if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){ std::string s = boost::algorithm::join(action, " "); - XBT_VERB("%s %f", s.c_str(), smpi_process()->simulated_elapsed()-clock); + XBT_VERB("%s %f", s.c_str(), smpi_process()->simulated_elapsed() - clock); } } -static std::vector* get_reqq_self() +/* Helper functions */ +static double parse_double(const std::string& string) { - return reqq.at(Actor::self()->getPid()); + try { + return boost::lexical_cast(string); + } catch (boost::bad_lexical_cast const&) { + throw std::invalid_argument("not a double: " + string); + } } -static void set_reqq_self(std::vector *mpi_request) +template static T parse_integer(const std::string& string) { - reqq.insert({Actor::self()->getPid(), mpi_request}); + 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); } -/* Helper function */ -static double parse_double(std::string string) +static int parse_root(const simgrid::xbt::ReplayAction& action, unsigned i) { - return xbt_str_parse_double(string.c_str(), "%s is not a double"); + return i < action.size() ? std::stoi(action[i]) : 0; } -namespace simgrid { -namespace smpi { - -namespace Replay { -class ActionArgParser { -public: - virtual void parse(simgrid::xbt::ReplayAction& action){}; -}; +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; +} -template class ReplayAction { - protected: - const std::string name; - T args; +namespace simgrid::smpi::replay { +MPI_Datatype MPI_DEFAULT_TYPE; - /* - * Used to compute the duration of this action. - */ - double start_time; +class RequestStorage { +private: + using req_key_t = std::tuple; + using req_storage_t = std::unordered_map, hash_tuple::hash>>; - int my_proc_id; + req_storage_t store; public: - explicit ReplayAction(std::string name) - : name(name), start_time(smpi_process()->simulated_elapsed()), my_proc_id(simgrid::s4u::Actor::self()->getPid()) + RequestStorage() = default; + size_t size() const { return store.size(); } + + req_storage_t& get_store() { return store; } + + void get_requests(std::vector& vec) const { + for (auto const& [_, reqs] : store) { + aid_t my_proc_id = simgrid::s4u::this_actor::get_pid(); + 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"); + } + } + } } - virtual void execute(simgrid::xbt::ReplayAction& action) + MPI_Request pop(int src, int dst, int tag) { - args.parse(action); - kernel(action); - log_timed_action(action, start_time); + 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; } - virtual void kernel(simgrid::xbt::ReplayAction& action) = 0; -}; - -class WaitAction : public ReplayAction { -public: - WaitAction() : ReplayAction("Wait") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void add(MPI_Request req) { - CHECK_ACTION_PARAMS(action, 0, 0) - MPI_Status status; - - std::string s = boost::algorithm::join(action, " "); - xbt_assert(get_reqq_self()->size(), "action wait not preceded by any irecv or isend: %s", s.c_str()); - MPI_Request request = get_reqq_self()->back(); - get_reqq_self()->pop_back(); - - if (request == nullptr) { - /* Assume that the trace is well formed, meaning the comm might have been caught by a MPI_test. Then just - * return.*/ - return; + 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); } + } - int rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1; - - MPI_Group group = request->comm()->group(); - int src_traced = group->rank(request->src()); - int dst_traced = group->rank(request->dst()); - bool is_wait_for_receive = (request->flags() & RECV); - // TODO: Here we take the rank while we normally take the process id (look for my_proc_id) - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::NoOpTIData("wait")); - - Request::wait(&request, &status); - - TRACE_smpi_comm_out(rank); - if (is_wait_for_receive) - TRACE_smpi_recv(src_traced, dst_traced, 0); + /* 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); } }; -static void action_init(simgrid::xbt::ReplayAction& action) +void WaitTestParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - XBT_DEBUG("Initialize the counters"); - CHECK_ACTION_PARAMS(action, 0, 1) - if (action.size() > 2) - MPI_DEFAULT_TYPE = MPI_DOUBLE; // default MPE datatype - else - MPI_DEFAULT_TYPE = MPI_BYTE; // default TAU datatype - - /* start a simulated timer */ - smpi_process()->simulated_start(); - /*initialize the number of active processes */ - active_processes = smpi_process_count(); - - set_reqq_self(new std::vector); + CHECK_ACTION_PARAMS(action, 3, 0) + src = std::stoi(action[2]); + dst = std::stoi(action[3]); + tag = std::stoi(action[4]); } -static void action_finalize(simgrid::xbt::ReplayAction& action) +void SendOrRecvParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - /* Nothing to do */ + CHECK_ACTION_PARAMS(action, 3, 1) + partner = std::stoi(action[2]); + tag = std::stoi(action[3]); + size = parse_integer(action[4]); + datatype1 = parse_datatype(action, 5); } -static void action_comm_size(simgrid::xbt::ReplayAction& action) +void ComputeParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - communicator_size = parse_double(action[2]); - log_timed_action (action, smpi_process()->simulated_elapsed()); + CHECK_ACTION_PARAMS(action, 1, 0) + flops = parse_double(action[2]); } -static void action_comm_split(simgrid::xbt::ReplayAction& action) +void SleepParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - log_timed_action (action, smpi_process()->simulated_elapsed()); + CHECK_ACTION_PARAMS(action, 1, 0) + time = parse_double(action[2]); } -static void action_comm_dup(simgrid::xbt::ReplayAction& action) +void LocationParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - log_timed_action (action, smpi_process()->simulated_elapsed()); + CHECK_ACTION_PARAMS(action, 2, 0) + filename = action[2]; + line = std::stoi(action[3]); } -static void action_compute(simgrid::xbt::ReplayAction& action) +void SendRecvParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - CHECK_ACTION_PARAMS(action, 1, 0) - double clock = smpi_process()->simulated_elapsed(); - double flops= parse_double(action[2]); - int my_proc_id = Actor::self()->getPid(); - - TRACE_smpi_computing_in(my_proc_id, flops); - smpi_execute_flops(flops); - TRACE_smpi_computing_out(my_proc_id); - - log_timed_action (action, clock); + 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); } -static void action_send(simgrid::xbt::ReplayAction& action) +void BcastArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - CHECK_ACTION_PARAMS(action, 2, 1) - int to = std::stoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; - - int my_proc_id = Actor::self()->getPid(); - int dst_traced = MPI_COMM_WORLD->group()->actor(to)->getPid(); - - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("send", to, size, Datatype::encode(MPI_CURRENT_TYPE))); - if (not TRACE_smpi_view_internals()) - TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, 0, size * MPI_CURRENT_TYPE->size()); - - Request::send(nullptr, size, MPI_CURRENT_TYPE, to , 0, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(my_proc_id); + CHECK_ACTION_PARAMS(action, 1, 2) + size = parse_integer(action[2]); + root = parse_root(action, 3); + datatype1 = parse_datatype(action, 4); +} - log_timed_action(action, clock); +void ReduceArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) +{ + CHECK_ACTION_PARAMS(action, 2, 2) + comm_size = parse_integer(action[2]); + comp_size = parse_double(action[3]); + root = parse_root(action, 4); + datatype1 = parse_datatype(action, 5); } -static void action_Isend(simgrid::xbt::ReplayAction& action) +void AllReduceArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { CHECK_ACTION_PARAMS(action, 2, 1) - int to = std::stoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; - - int my_proc_id = Actor::self()->getPid(); - int dst_traced = MPI_COMM_WORLD->group()->actor(to)->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("Isend", to, size, Datatype::encode(MPI_CURRENT_TYPE))); - if (not TRACE_smpi_view_internals()) - TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, 0, size * MPI_CURRENT_TYPE->size()); - - MPI_Request request = Request::isend(nullptr, size, MPI_CURRENT_TYPE, to, 0, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(my_proc_id); - - get_reqq_self()->push_back(request); - - log_timed_action (action, clock); + comm_size = parse_integer(action[2]); + comp_size = parse_double(action[3]); + datatype1 = parse_datatype(action, 4); } -static void action_recv(simgrid::xbt::ReplayAction& action) +void AllToAllArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { CHECK_ACTION_PARAMS(action, 2, 1) - int from = std::stoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - MPI_Status status; - - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; + comm_size = MPI_COMM_WORLD->size(); + send_size = parse_integer(action[2]); + recv_size = parse_integer(action[3]); + datatype1 = parse_datatype(action, 4); + datatype2 = parse_datatype(action, 5); +} - int my_proc_id = Actor::self()->getPid(); - int src_traced = MPI_COMM_WORLD->group()->actor(from)->getPid(); +void GatherArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string& name) +{ + /* The structure of the gather action for the rank 0 (total 4 processes) is the following: + 0 gather 68 68 0 0 0 + where: + 1) 68 is the sendcounts + 2) 68 is the recvcounts + 3) 0 is the root node + 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() + 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() + */ + CHECK_ACTION_PARAMS(action, 2, 3) + comm_size = MPI_COMM_WORLD->size(); + send_size = parse_integer(action[2]); + recv_size = parse_integer(action[3]); + + if (name == "gather") { + root = parse_root(action, 4); + datatype1 = parse_datatype(action, 5); + datatype2 = parse_datatype(action, 6); + } else { + root = 0; + datatype1 = parse_datatype(action, 4); + datatype2 = parse_datatype(action, 5); + } +} - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("recv", from, size, Datatype::encode(MPI_CURRENT_TYPE))); +void GatherVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string& name) +{ + /* The structure of the gatherv action for the rank 0 (total 4 processes) is the following: + 0 gather 68 68 10 10 10 0 0 0 + where: + 1) 68 is the sendcount + 2) 68 10 10 10 is the recvcounts + 3) 0 is the root node + 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) + send_size = parse_integer(action[2]); + disps = std::vector(comm_size, 0); + recvcounts = std::make_shared>(comm_size); + + if (name == "gatherv") { + root = parse_root(action, 3 + comm_size); + datatype1 = parse_datatype(action, 4 + comm_size); + datatype2 = parse_datatype(action, 5 + comm_size); + } else { + 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 + 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 { + // no disp specified, maybe only datatype, + datatype1 = parse_datatype(action, 3 + comm_size); + datatype2 = parse_datatype(action, 4 + comm_size); + } - //unknown size from the receiver point of view - if (size <= 0.0) { - Request::probe(from, 0, MPI_COMM_WORLD, &status); - size=status.count; + if (disp_index != 0) { + xbt_assert(disp_index + comm_size <= action.size()); + for (unsigned i = 0; i < comm_size; i++) + disps[i] = std::stoi(action[disp_index + i]); + } } - Request::recv(nullptr, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD, &status); - - TRACE_smpi_comm_out(my_proc_id); - if (not TRACE_smpi_view_internals()) { - TRACE_smpi_recv(src_traced, my_proc_id, 0); + for (unsigned int i = 0; i < comm_size; i++) { + (*recvcounts)[i] = std::stoi(action[i + 3]); } + recv_size_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); +} - log_timed_action (action, clock); +void ScatterArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) +{ + /* The structure of the scatter action for the rank 0 (total 4 processes) is the following: + 0 gather 68 68 0 0 0 + where: + 1) 68 is the sendcounts + 2) 68 is the recvcounts + 3) 0 is the root node + 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(action[2]); + recv_size = parse_integer(action[3]); + root = parse_root(action, 4); + datatype1 = parse_datatype(action, 5); + datatype2 = parse_datatype(action, 6); } -static void action_Irecv(simgrid::xbt::ReplayAction& action) +void ScatterVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - CHECK_ACTION_PARAMS(action, 2, 1) - int from = std::stoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); + /* The structure of the scatterv action for the rank 0 (total 4 processes) is the following: + 0 gather 68 10 10 10 68 0 0 0 + where: + 1) 68 10 10 10 is the sendcounts + 2) 68 is the recvcount + 3) 0 is the root node + 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(action[2 + comm_size]); + disps = std::vector(comm_size, 0); + sendcounts = std::make_shared>(comm_size); - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; + root = parse_root(action, 3 + comm_size); + datatype1 = parse_datatype(action, 4 + comm_size); + datatype2 = parse_datatype(action, 5 + comm_size); - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("Irecv", from, size, Datatype::encode(MPI_CURRENT_TYPE))); - MPI_Status status; - //unknow size from the receiver pov - if (size <= 0.0) { - Request::probe(from, 0, MPI_COMM_WORLD, &status); - size = status.count; + for (unsigned int i = 0; i < comm_size; i++) { + (*sendcounts)[i] = std::stoi(action[i + 2]); } - - MPI_Request request = Request::irecv(nullptr, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(my_proc_id); - get_reqq_self()->push_back(request); - - log_timed_action (action, clock); + send_size_sum = std::accumulate(sendcounts->begin(), sendcounts->end(), 0); } -static void action_test(simgrid::xbt::ReplayAction& action) +void ReduceScatterArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - CHECK_ACTION_PARAMS(action, 0, 0) - double clock = smpi_process()->simulated_elapsed(); - MPI_Status status; - - MPI_Request request = get_reqq_self()->back(); - get_reqq_self()->pop_back(); - //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!=nullptr){ - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_testing_in(my_proc_id); - - int flag = Request::test(&request, &status); - - XBT_DEBUG("MPI_Test result: %d", flag); - /* push back request in vector to be caught by a subsequent wait. if the test did succeed, the request is now nullptr.*/ - get_reqq_self()->push_back(request); + /* The structure of the reducescatter action for the rank 0 (total 4 processes) is the following: + 0 reducescatter 275427 275427 275427 204020 11346849 0 + where: + 1) The first four values after the name of the action declare the recvcounts array + 2) The value 11346849 is the amount of instructions + 3) The last value corresponds to the datatype, see simgrid::smpi::Datatype::decode(). + */ + comm_size = MPI_COMM_WORLD->size(); + CHECK_ACTION_PARAMS(action, comm_size + 1, 1) + comp_size = parse_double(action[2 + comm_size]); + recvcounts = std::make_shared>(comm_size); + datatype1 = parse_datatype(action, 3 + comm_size); - TRACE_smpi_testing_out(my_proc_id); + for (unsigned int i = 0; i < comm_size; i++) { + (*recvcounts)[i]= std::stoi(action[i + 2]); } - log_timed_action (action, clock); + recv_size_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); } -static void action_wait(simgrid::xbt::ReplayAction& action) +void ScanArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - Replay::WaitAction().execute(action); + CHECK_ACTION_PARAMS(action, 2, 1) + size = parse_integer(action[2]); + comp_size = parse_double(action[3]); + datatype1 = parse_datatype(action, 4); } -static void action_waitall(simgrid::xbt::ReplayAction& action) +void AllToAllVArgParser::parse(simgrid::xbt::ReplayAction& action, const std::string&) { - CHECK_ACTION_PARAMS(action, 0, 0) - double clock = smpi_process()->simulated_elapsed(); - const unsigned int count_requests = get_reqq_self()->size(); - - if (count_requests>0) { - MPI_Status status[count_requests]; - - int my_proc_id_traced = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id_traced, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("waitAll", -1, count_requests, "")); - int recvs_snd[count_requests]; - int recvs_rcv[count_requests]; - for (unsigned int i = 0; i < count_requests; i++) { - const auto& req = (*get_reqq_self())[i]; - if (req && (req->flags() & RECV)) { - recvs_snd[i] = req->src(); - recvs_rcv[i] = req->dst(); - } else - recvs_snd[i] = -100; - } - Request::waitall(count_requests, &(*get_reqq_self())[0], status); - - for (unsigned i = 0; i < count_requests; i++) { - if (recvs_snd[i]!=-100) - TRACE_smpi_recv(recvs_snd[i], recvs_rcv[i],0); - } - TRACE_smpi_comm_out(my_proc_id_traced); + /* The structure of the alltoallv action for the rank 0 (total 4 processes) is the following: + 0 alltoallv 100 1 7 10 12 100 1 70 10 5 + where: + 1) 100 is the size of the send buffer *sizeof(int), + 2) 1 7 10 12 is the sendcounts array + 3) 100*sizeof(int) is the size of the receiver buffer + 4) 1 70 10 5 is the recvcounts array + */ + comm_size = MPI_COMM_WORLD->size(); + CHECK_ACTION_PARAMS(action, 2 * comm_size + 2, 2) + sendcounts = std::make_shared>(comm_size); + recvcounts = std::make_shared>(comm_size); + senddisps = std::vector(comm_size, 0); + recvdisps = std::vector(comm_size, 0); + + datatype1 = parse_datatype(action, 4 + 2 * comm_size); + datatype2 = parse_datatype(action, 5 + 2 * 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]); } - log_timed_action (action, clock); -} - -static void action_barrier(simgrid::xbt::ReplayAction& action) -{ - double clock = smpi_process()->simulated_elapsed(); - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::NoOpTIData("barrier")); - - Colls::barrier(MPI_COMM_WORLD); - - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + send_size_sum = std::accumulate(sendcounts->begin(), sendcounts->end(), 0); + recv_size_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); } -static void action_bcast(simgrid::xbt::ReplayAction& action) +void WaitAction::kernel(simgrid::xbt::ReplayAction& action) { - CHECK_ACTION_PARAMS(action, 1, 2) - double size = parse_double(action[2]); - double clock = smpi_process()->simulated_elapsed(); - int root = (action.size() > 3) ? std::stoi(action[3]) : 0; - /* Initialize MPI_CURRENT_TYPE in order to decrease the number of the checks */ - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; + 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.pop(args.src, args.dst, args.tag); - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->actor(root)->getPid(), -1.0, size, - -1, Datatype::encode(MPI_CURRENT_TYPE), "")); + 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 + * return.*/ + return; + } - void *sendbuf = smpi_get_tmp_sendbuffer(size* MPI_CURRENT_TYPE->size()); + // 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); - Colls::bcast(sendbuf, size, MPI_CURRENT_TYPE, root, MPI_COMM_WORLD); + TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::WaitTIData("wait", args.src, args.dst, args.tag)); - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + 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); } -static void action_reduce(simgrid::xbt::ReplayAction& action) +void SendAction::kernel(simgrid::xbt::ReplayAction&) { - CHECK_ACTION_PARAMS(action, 2, 2) - double comm_size = parse_double(action[2]); - double comp_size = parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - int root = (action.size() > 4) ? std::stoi(action[4]) : 0; + const SendOrRecvParser& args = get_args(); + aid_t dst_traced = MPI_COMM_WORLD->group()->actor(args.partner); - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 5) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE; - - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::CollTIData("reduce", MPI_COMM_WORLD->group()->actor(root)->getPid(), comp_size, - comm_size, -1, Datatype::encode(MPI_CURRENT_TYPE), "")); + TRACE_smpi_comm_in( + get_pid(), __func__, + new simgrid::instr::Pt2PtTIData(get_name(), args.partner, args.size, args.tag, Datatype::encode(args.datatype1))); + if (not TRACE_smpi_view_internals()) + TRACE_smpi_send(get_pid(), get_pid(), dst_traced, args.tag, args.size * args.datatype1->size()); - void *recvbuf = smpi_get_tmp_sendbuffer(comm_size* MPI_CURRENT_TYPE->size()); - void *sendbuf = smpi_get_tmp_sendbuffer(comm_size* MPI_CURRENT_TYPE->size()); - Colls::reduce(sendbuf, recvbuf, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, root, MPI_COMM_WORLD); - smpi_execute_flops(comp_size); + if (get_name() == "send") { + Request::send(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD); + } else if (get_name() == "isend") { + MPI_Request request = Request::isend(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD); + req_storage.add(request); + } else { + xbt_die("Don't know this action, %s", get_name().c_str()); + } - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + TRACE_smpi_comm_out(get_pid()); } -static void action_allReduce(simgrid::xbt::ReplayAction& action) +void RecvAction::kernel(simgrid::xbt::ReplayAction&) { - CHECK_ACTION_PARAMS(action, 2, 1) - double comm_size = parse_double(action[2]); - double comp_size = parse_double(action[3]); - - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 4) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE; + 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))); - double clock = smpi_process()->simulated_elapsed(); - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::CollTIData("allReduce", -1, comp_size, comm_size, -1, - Datatype::encode(MPI_CURRENT_TYPE), "")); + MPI_Status status; + // unknown size from the receiver point of view + 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 *recvbuf = smpi_get_tmp_recvbuffer(comm_size* MPI_CURRENT_TYPE->size()); - void *sendbuf = smpi_get_tmp_sendbuffer(comm_size* MPI_CURRENT_TYPE->size()); - Colls::allreduce(sendbuf, recvbuf, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, MPI_COMM_WORLD); - smpi_execute_flops(comp_size); + bool is_recv = false; // Help analyzers understanding that status is not used uninitialized + if (get_name() == "recv") { + is_recv = true; + Request::recv(nullptr, arg_size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD, &status); + } else if (get_name() == "irecv") { + MPI_Request request = Request::irecv(nullptr, arg_size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD); + req_storage.add(request); + } else { + THROW_IMPOSSIBLE; + } - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + TRACE_smpi_comm_out(get_pid()); + if (is_recv && not TRACE_smpi_view_internals()) { + aid_t src_traced = MPI_COMM_WORLD->group()->actor(status.MPI_SOURCE); + TRACE_smpi_recv(src_traced, get_pid(), args.tag); + } } -static void action_allToAll(simgrid::xbt::ReplayAction& action) +void SendRecvAction::kernel(simgrid::xbt::ReplayAction&) { - CHECK_ACTION_PARAMS(action, 2, 2) //two mandatory (send and recv volumes) and two optional (corresponding datatypes) - double clock = smpi_process()->simulated_elapsed(); - unsigned long comm_size = MPI_COMM_WORLD->size(); - int send_size = parse_double(action[2]); - int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE}; - MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; - - void *send = smpi_get_tmp_sendbuffer(send_size*comm_size* MPI_CURRENT_TYPE->size()); - void *recv = smpi_get_tmp_recvbuffer(recv_size*comm_size* MPI_CURRENT_TYPE2->size()); - - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::CollTIData("allToAll", -1, -1.0, send_size, recv_size, - Datatype::encode(MPI_CURRENT_TYPE), - Datatype::encode(MPI_CURRENT_TYPE2))); - - Colls::alltoall(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); + 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); - log_timed_action (action, clock); + XBT_DEBUG("Exits SendRecv"); } -static void action_gather(simgrid::xbt::ReplayAction& action) +void ComputeAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the gather action for the rank 0 (total 4 processes) is the following: - 0 gather 68 68 0 0 0 - where: - 1) 68 is the sendcounts - 2) 68 is the recvcounts - 3) 0 is the root node - 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() - 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() - */ - CHECK_ACTION_PARAMS(action, 2, 3) - double clock = smpi_process()->simulated_elapsed(); - unsigned long comm_size = MPI_COMM_WORLD->size(); - int send_size = parse_double(action[2]); - int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; - MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[6]) : MPI_DEFAULT_TYPE}; - - void *send = smpi_get_tmp_sendbuffer(send_size* MPI_CURRENT_TYPE->size()); - void *recv = nullptr; - int root = (action.size() > 4) ? std::stoi(action[4]) : 0; - int rank = MPI_COMM_WORLD->rank(); - - if(rank==root) - recv = smpi_get_tmp_recvbuffer(recv_size*comm_size* MPI_CURRENT_TYPE2->size()); - - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("gather", root, -1.0, send_size, recv_size, - Datatype::encode(MPI_CURRENT_TYPE), - Datatype::encode(MPI_CURRENT_TYPE2))); - - Colls::gather(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(Actor::self()->getPid()); - log_timed_action (action, clock); + const ComputeParser& args = get_args(); + if (smpi_cfg_simulate_computation()) { + smpi_execute_flops(args.flops/smpi_adjust_comp_speed()); + } } -static void action_scatter(simgrid::xbt::ReplayAction& action) +void SleepAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the scatter action for the rank 0 (total 4 processes) is the following: - 0 gather 68 68 0 0 0 - where: - 1) 68 is the sendcounts - 2) 68 is the recvcounts - 3) 0 is the root node - 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() - 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() - */ - CHECK_ACTION_PARAMS(action, 2, 3) - double clock = smpi_process()->simulated_elapsed(); - unsigned long comm_size = MPI_COMM_WORLD->size(); - int send_size = parse_double(action[2]); - int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; - MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 6) ? simgrid::smpi::Datatype::decode(action[6]) : MPI_DEFAULT_TYPE}; - - void* send = smpi_get_tmp_sendbuffer(send_size * MPI_CURRENT_TYPE->size()); - void* recv = nullptr; - int root = (action.size() > 4) ? std::stoi(action[4]) : 0; - int rank = MPI_COMM_WORLD->rank(); - - if (rank == root) - recv = smpi_get_tmp_recvbuffer(recv_size * comm_size * MPI_CURRENT_TYPE2->size()); - - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("gather", root, -1.0, send_size, recv_size, - Datatype::encode(MPI_CURRENT_TYPE), - Datatype::encode(MPI_CURRENT_TYPE2))); - - Colls::scatter(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); + const SleepParser& args = get_args(); + XBT_DEBUG("Sleep for: %lf secs", 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(pid); +} - TRACE_smpi_comm_out(Actor::self()->getPid()); - log_timed_action(action, clock); +void LocationAction::kernel(simgrid::xbt::ReplayAction&) +{ + const LocationParser& args = get_args(); + smpi_trace_set_call_location(args.filename.c_str(), args.line, "replay_action"); } -static void action_gatherv(simgrid::xbt::ReplayAction& action) +void TestAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the gatherv action for the rank 0 (total 4 processes) is the following: - 0 gather 68 68 10 10 10 0 0 0 - where: - 1) 68 is the sendcount - 2) 68 10 10 10 is the recvcounts - 3) 0 is the root node - 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() - 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() - */ - double clock = smpi_process()->simulated_elapsed(); - unsigned long comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size+1, 2) - int send_size = parse_double(action[2]); - std::vector disps(comm_size, 0); - std::shared_ptr> recvcounts(new std::vector(comm_size)); - - MPI_Datatype MPI_CURRENT_TYPE = - (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[4 + comm_size]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{ - (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[5 + comm_size]) : MPI_DEFAULT_TYPE}; - - void *send = smpi_get_tmp_sendbuffer(send_size* MPI_CURRENT_TYPE->size()); - void *recv = nullptr; - for (unsigned int i = 0; i < comm_size; i++) { - (*recvcounts)[i] = std::stoi(action[i + 3]); - } - int recv_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); + const WaitTestParser& args = get_args(); + 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::WaitTIData("test", args.src, args.dst, args.tag)); - int root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; - int rank = MPI_COMM_WORLD->rank(); + MPI_Status status; + int flag = 0; + Request::test(&request, &status, &flag); - if(rank==root) - recv = smpi_get_tmp_recvbuffer(recv_sum* MPI_CURRENT_TYPE2->size()); + XBT_DEBUG("MPI_Test result: %d", flag); + /* push back request in vector to be caught by a subsequent wait. if the test did succeed, the request is now + * nullptr.*/ + if (request == MPI_REQUEST_NULL) + req_storage.addNullRequest(args.src, args.dst, args.tag); + else + req_storage.add(request); + + TRACE_smpi_comm_out(get_pid()); + } +} - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData( - "gatherV", root, send_size, nullptr, -1, recvcounts, - Datatype::encode(MPI_CURRENT_TYPE), Datatype::encode(MPI_CURRENT_TYPE2))); +void InitAction::kernel(simgrid::xbt::ReplayAction& action) +{ + CHECK_ACTION_PARAMS(action, 0, 1) + MPI_DEFAULT_TYPE = (action.size() > 2) ? MPI_DOUBLE // default MPE datatype + : MPI_BYTE; // default TAU datatype - Colls::gatherv(send, send_size, MPI_CURRENT_TYPE, recv, recvcounts->data(), disps.data(), MPI_CURRENT_TYPE2, root, - MPI_COMM_WORLD); + /* start a simulated timer */ + smpi_process()->simulated_start(); +} - TRACE_smpi_comm_out(Actor::self()->getPid()); - log_timed_action (action, clock); +void CommunicatorAction::kernel(simgrid::xbt::ReplayAction&) +{ + /* nothing to do */ } -static void action_scatterv(simgrid::xbt::ReplayAction& action) +void WaitAllAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the scatterv action for the rank 0 (total 4 processes) is the following: - 0 gather 68 10 10 10 68 0 0 0 - where: - 1) 68 10 10 10 is the sendcounts - 2) 68 is the recvcount - 3) 0 is the root node - 4) 0 is the send datatype id, see simgrid::smpi::Datatype::decode() - 5) 0 is the recv datatype id, see simgrid::smpi::Datatype::decode() - */ - double clock = smpi_process()->simulated_elapsed(); - unsigned long comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size + 1, 2) - int recv_size = parse_double(action[2 + comm_size]); - std::vector disps(comm_size, 0); - std::shared_ptr> sendcounts(new std::vector(comm_size)); + 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()); + } + } + Request::waitall(count_requests, &(reqs.data())[0], MPI_STATUSES_IGNORE); + req_storage.get_store().clear(); - MPI_Datatype MPI_CURRENT_TYPE = - (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[4 + comm_size]) : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{ - (action.size() > 5 + comm_size) ? simgrid::smpi::Datatype::decode(action[5 + comm_size]) : MPI_DEFAULT_TYPE}; + for (MPI_Request& req : reqs) + if (req != MPI_REQUEST_NULL) + Request::unref(&req); - void* send = nullptr; - void* recv = smpi_get_tmp_recvbuffer(recv_size * MPI_CURRENT_TYPE->size()); - for (unsigned int i = 0; i < comm_size; i++) { - (*sendcounts)[i] = std::stoi(action[i + 2]); + for (auto const& [src, dst] : sender_receiver) { + TRACE_smpi_recv(src, dst, 0); + } + TRACE_smpi_comm_out(get_pid()); } - int send_sum = std::accumulate(sendcounts->begin(), sendcounts->end(), 0); - - int root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; - int rank = MPI_COMM_WORLD->rank(); +} - if (rank == root) - send = smpi_get_tmp_sendbuffer(send_sum * MPI_CURRENT_TYPE2->size()); +void BarrierAction::kernel(simgrid::xbt::ReplayAction&) +{ + TRACE_smpi_comm_in(get_pid(), __func__, new simgrid::instr::NoOpTIData("barrier")); + colls::barrier(MPI_COMM_WORLD); + TRACE_smpi_comm_out(get_pid()); +} - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData("gatherV", root, -1, sendcounts, recv_size, - nullptr, Datatype::encode(MPI_CURRENT_TYPE), - Datatype::encode(MPI_CURRENT_TYPE2))); +void BcastAction::kernel(simgrid::xbt::ReplayAction&) +{ + const BcastArgParser& args = get_args(); + TRACE_smpi_comm_in(get_pid(), "action_bcast", + new simgrid::instr::CollTIData("bcast", args.root, -1.0, args.size, + 0, Datatype::encode(args.datatype1), "")); - Colls::scatterv(send, sendcounts->data(), disps.data(), MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, - MPI_COMM_WORLD); + colls::bcast(send_buffer(args.size * args.datatype1->size()), args.size, args.datatype1, args.root, MPI_COMM_WORLD); - TRACE_smpi_comm_out(Actor::self()->getPid()); - log_timed_action(action, clock); + TRACE_smpi_comm_out(get_pid()); } -static void action_reducescatter(simgrid::xbt::ReplayAction& action) +void ReduceAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the reducescatter action for the rank 0 (total 4 processes) is the following: - 0 reduceScatter 275427 275427 275427 204020 11346849 0 - where: - 1) The first four values after the name of the action declare the recvcounts array - 2) The value 11346849 is the amount of instructions - 3) The last value corresponds to the datatype, see simgrid::smpi::Datatype::decode(). - */ - double clock = smpi_process()->simulated_elapsed(); - unsigned long comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size+1, 1) - int comp_size = parse_double(action[2+comm_size]); - int my_proc_id = Actor::self()->getPid(); - std::shared_ptr> recvcounts(new std::vector); - MPI_Datatype MPI_CURRENT_TYPE = - (action.size() > 3 + comm_size) ? simgrid::smpi::Datatype::decode(action[3 + comm_size]) : MPI_DEFAULT_TYPE; + const ReduceArgParser& args = get_args(); + TRACE_smpi_comm_in(get_pid(), "action_reduce", + new simgrid::instr::CollTIData("reduce", args.root, args.comp_size, + args.comm_size, 0, Datatype::encode(args.datatype1), "")); - for (unsigned int i = 0; i < comm_size; i++) { - recvcounts->push_back(std::stoi(action[i + 2])); - } - int size{std::accumulate(recvcounts->begin(), recvcounts->end(), 0)}; + 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) + simgrid::s4u::this_actor::exec_init(args.comp_size) + ->set_name("computation") + ->start() + ->wait(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::VarCollTIData("reduceScatter", -1, 0, nullptr, -1, recvcounts, - std::to_string(comp_size), /* ugly hack to print comp_size */ - Datatype::encode(MPI_CURRENT_TYPE))); + TRACE_smpi_comm_out(get_pid()); +} - void *sendbuf = smpi_get_tmp_sendbuffer(size* MPI_CURRENT_TYPE->size()); - void *recvbuf = smpi_get_tmp_recvbuffer(size* MPI_CURRENT_TYPE->size()); +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, 0, + Datatype::encode(args.datatype1), "")); - Colls::reduce_scatter(sendbuf, recvbuf, recvcounts->data(), MPI_CURRENT_TYPE, MPI_OP_NULL, MPI_COMM_WORLD); - smpi_execute_flops(comp_size); + 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) + simgrid::s4u::this_actor::exec_init(args.comp_size) + ->set_name("computation") + ->start() + ->wait(); - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + TRACE_smpi_comm_out(get_pid()); } -static void action_allgather(simgrid::xbt::ReplayAction& action) +void AllToAllAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the allgather action for the rank 0 (total 4 processes) is the following: - 0 allGather 275427 275427 - where: - 1) 275427 is the sendcount - 2) 275427 is the recvcount - 3) No more values mean that the datatype for sent and receive buffer is the default one, see - simgrid::smpi::Datatype::decode(). - */ - double clock = smpi_process()->simulated_elapsed(); - - CHECK_ACTION_PARAMS(action, 2, 2) - int sendcount = std::stoi(action[2]); - int recvcount = std::stoi(action[3]); - - MPI_Datatype MPI_CURRENT_TYPE{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[4]) : MPI_DEFAULT_TYPE}; - MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 5) ? simgrid::smpi::Datatype::decode(action[5]) : MPI_DEFAULT_TYPE}; + const AllToAllArgParser& args = get_args(); + TRACE_smpi_comm_in(get_pid(), "action_alltoall", + new simgrid::instr::CollTIData("alltoall", -1, -1.0, args.send_size, args.recv_size, + Datatype::encode(args.datatype1), + Datatype::encode(args.datatype2))); - void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* MPI_CURRENT_TYPE->size()); - void *recvbuf = smpi_get_tmp_recvbuffer(recvcount* MPI_CURRENT_TYPE2->size()); + 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); - int my_proc_id = Actor::self()->getPid(); + TRACE_smpi_comm_out(get_pid()); +} - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::CollTIData("allGather", -1, -1.0, sendcount, recvcount, - Datatype::encode(MPI_CURRENT_TYPE), - Datatype::encode(MPI_CURRENT_TYPE2))); +void GatherAction::kernel(simgrid::xbt::ReplayAction&) +{ + const GatherArgParser& args = get_args(); + TRACE_smpi_comm_in(get_pid(), get_name().c_str(), + new simgrid::instr::CollTIData(get_name(), (get_name() == "gather") ? args.root : -1, -1.0, + args.send_size, args.recv_size, Datatype::encode(args.datatype1), + Datatype::encode(args.datatype2))); - Colls::allgather(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcount, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); + 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.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, + recv_buffer(args.recv_size * args.datatype2->size()), args.recv_size, args.datatype2, + MPI_COMM_WORLD); - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + TRACE_smpi_comm_out(get_pid()); } -static void action_allgatherv(simgrid::xbt::ReplayAction& action) +void GatherVAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the allgatherv action for the rank 0 (total 4 processes) is the following: - 0 allGatherV 275427 275427 275427 275427 204020 - where: - 1) 275427 is the sendcount - 2) The next four elements declare the recvcounts array - 3) No more values mean that the datatype for sent and receive buffer is the default one, see - simgrid::smpi::Datatype::decode(). - */ - double clock = smpi_process()->simulated_elapsed(); - - unsigned long comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size+1, 2) - int sendcount = std::stoi(action[2]); - std::shared_ptr> recvcounts(new std::vector(comm_size)); - std::vector disps(comm_size, 0); - - int datatype_index = 0, disp_index = 0; - if (action.size() > 3 + 2 * comm_size) { /* datatype + disp are specified */ - datatype_index = 3 + comm_size; - disp_index = datatype_index + 1; - } else if (action.size() > 3 + 2 * comm_size) { /* disps specified; datatype is not specified; use the default one */ - datatype_index = -1; - disp_index = 3 + comm_size; - } else if (action.size() > 3 + comm_size) { /* only datatype, no disp specified */ - datatype_index = 3 + comm_size; - } - - if (disp_index != 0) { - for (unsigned int i = 0; i < comm_size; i++) - disps[i] = std::stoi(action[disp_index + i]); + int rank = MPI_COMM_WORLD->rank(); + 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, + args.recvcounts, Datatype::encode(args.datatype1), Datatype::encode(args.datatype2))); + + if (get_name() == "gatherv") { + colls::gatherv(send_buffer(args.send_size * args.datatype1->size()), args.send_size, args.datatype1, + (rank == args.root) ? recv_buffer(args.recv_size_sum * args.datatype2->size()) : nullptr, + args.recvcounts->data(), args.disps.data(), args.datatype2, args.root, MPI_COMM_WORLD); + } else { + colls::allgatherv(send_buffer(args.send_size * args.datatype1->size()), args.send_size, args.datatype1, + recv_buffer(args.recv_size_sum * args.datatype2->size()), args.recvcounts->data(), + args.disps.data(), args.datatype2, MPI_COMM_WORLD); } - MPI_Datatype MPI_CURRENT_TYPE{(datatype_index > 0) ? simgrid::smpi::Datatype::decode(action[datatype_index]) - : MPI_DEFAULT_TYPE}; - MPI_Datatype MPI_CURRENT_TYPE2{(datatype_index > 0) ? simgrid::smpi::Datatype::decode(action[datatype_index]) - : MPI_DEFAULT_TYPE}; + TRACE_smpi_comm_out(get_pid()); +} - void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* MPI_CURRENT_TYPE->size()); +void ScatterAction::kernel(simgrid::xbt::ReplayAction&) +{ + int rank = MPI_COMM_WORLD->rank(); + const ScatterArgParser& args = get_args(); + TRACE_smpi_comm_in(get_pid(), "action_scatter", + new simgrid::instr::CollTIData(get_name(), args.root, -1.0, args.send_size, args.recv_size, + Datatype::encode(args.datatype1), + Datatype::encode(args.datatype2))); - for (unsigned int i = 0; i < comm_size; i++) { - (*recvcounts)[i] = std::stoi(action[i + 3]); - } - int recv_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); - void *recvbuf = smpi_get_tmp_recvbuffer(recv_sum* MPI_CURRENT_TYPE2->size()); + colls::scatter(send_buffer(args.send_size * args.datatype1->size()), args.send_size, args.datatype1, + (rank == args.root) ? recv_buffer(args.recv_size * args.datatype2->size()) : nullptr, args.recv_size, + args.datatype2, args.root, MPI_COMM_WORLD); - int my_proc_id = Actor::self()->getPid(); + TRACE_smpi_comm_out(get_pid()); +} - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::VarCollTIData("allGatherV", -1, sendcount, nullptr, -1, recvcounts, - Datatype::encode(MPI_CURRENT_TYPE), - Datatype::encode(MPI_CURRENT_TYPE2))); +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, + nullptr, Datatype::encode(args.datatype1), + Datatype::encode(args.datatype2))); + + colls::scatterv((rank == args.root) ? send_buffer(args.send_size_sum * args.datatype1->size()) : nullptr, + args.sendcounts->data(), args.disps.data(), args.datatype1, + recv_buffer(args.recv_size * args.datatype2->size()), args.recv_size, args.datatype2, args.root, + MPI_COMM_WORLD); - Colls::allgatherv(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcounts->data(), disps.data(), MPI_CURRENT_TYPE2, - MPI_COMM_WORLD); + TRACE_smpi_comm_out(get_pid()); +} + +void ReduceScatterAction::kernel(simgrid::xbt::ReplayAction&) +{ + const ReduceScatterArgParser& args = get_args(); + TRACE_smpi_comm_in( + get_pid(), "action_reducescatter", + 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()); +} + +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); - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + 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()); } -static void action_allToAllv(simgrid::xbt::ReplayAction& action) +void AllToAllVAction::kernel(simgrid::xbt::ReplayAction&) { - /* The structure of the allToAllV action for the rank 0 (total 4 processes) is the following: - 0 allToAllV 100 1 7 10 12 100 1 70 10 5 - where: - 1) 100 is the size of the send buffer *sizeof(int), - 2) 1 7 10 12 is the sendcounts array - 3) 100*sizeof(int) is the size of the receiver buffer - 4) 1 70 10 5 is the recvcounts array - */ - double clock = smpi_process()->simulated_elapsed(); - - unsigned long comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, 2*comm_size+2, 2) - std::shared_ptr> sendcounts(new std::vector(comm_size)); - std::shared_ptr> recvcounts(new std::vector(comm_size)); - std::vector senddisps(comm_size, 0); - std::vector recvdisps(comm_size, 0); - - MPI_Datatype MPI_CURRENT_TYPE = (action.size() > 5 + 2 * comm_size) - ? simgrid::smpi::Datatype::decode(action[4 + 2 * comm_size]) - : MPI_DEFAULT_TYPE; - MPI_Datatype MPI_CURRENT_TYPE2{(action.size() > 5 + 2 * comm_size) - ? simgrid::smpi::Datatype::decode(action[5 + 2 * comm_size]) - : MPI_DEFAULT_TYPE}; - - int send_buf_size=parse_double(action[2]); - int recv_buf_size=parse_double(action[3+comm_size]); - int my_proc_id = Actor::self()->getPid(); - void *sendbuf = smpi_get_tmp_sendbuffer(send_buf_size* MPI_CURRENT_TYPE->size()); - void *recvbuf = smpi_get_tmp_recvbuffer(recv_buf_size* MPI_CURRENT_TYPE2->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]); - } - int send_size = std::accumulate(sendcounts->begin(), sendcounts->end(), 0); - int recv_size = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); - - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::VarCollTIData("allToAllV", -1, send_size, sendcounts, recv_size, recvcounts, - Datatype::encode(MPI_CURRENT_TYPE), - Datatype::encode(MPI_CURRENT_TYPE2))); + 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, + Datatype::encode(args.datatype1), Datatype::encode(args.datatype2))); - Colls::alltoallv(sendbuf, sendcounts->data(), senddisps.data(), MPI_CURRENT_TYPE, recvbuf, recvcounts->data(), - recvdisps.data(), MPI_CURRENT_TYPE, MPI_COMM_WORLD); + colls::alltoallv(send_buffer(args.send_buf_size * args.datatype1->size()), args.sendcounts->data(), + args.senddisps.data(), args.datatype1, recv_buffer(args.recv_buf_size * args.datatype2->size()), + args.recvcounts->data(), args.recvdisps.data(), args.datatype2, MPI_COMM_WORLD); - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); + TRACE_smpi_comm_out(get_pid()); } +} // namespace simgrid::smpi::replay -}} // namespace simgrid::smpi - +static std::unordered_map storage; /** @brief Only initialize the replay, don't do it for real */ -void smpi_replay_init(int* argc, char*** argv) +void smpi_replay_init(const char* instance_id, int rank, double start_delay_flops) { - simgrid::smpi::Process::init(argc, argv); + xbt_assert(not smpi_process()->initializing()); + + simgrid::s4u::Actor::self()->set_property("instance_id", instance_id); + simgrid::s4u::Actor::self()->set_property("rank", std::to_string(rank)); + simgrid::smpi::ActorExt::init(); + smpi_process()->mark_as_initialized(); smpi_process()->set_replaying(true); - int my_proc_id = Actor::self()->getPid(); - TRACE_smpi_init(my_proc_id); - TRACE_smpi_computing_init(my_proc_id); - TRACE_smpi_comm_in(my_proc_id, "smpi_replay_run_init", new simgrid::instr::NoOpTIData("init")); - TRACE_smpi_comm_out(my_proc_id); - xbt_replay_action_register("init", simgrid::smpi::action_init); - xbt_replay_action_register("finalize", simgrid::smpi::action_finalize); - xbt_replay_action_register("comm_size", simgrid::smpi::action_comm_size); - xbt_replay_action_register("comm_split", simgrid::smpi::action_comm_split); - xbt_replay_action_register("comm_dup", simgrid::smpi::action_comm_dup); - xbt_replay_action_register("send", simgrid::smpi::action_send); - xbt_replay_action_register("Isend", simgrid::smpi::action_Isend); - xbt_replay_action_register("recv", simgrid::smpi::action_recv); - xbt_replay_action_register("Irecv", simgrid::smpi::action_Irecv); - xbt_replay_action_register("test", simgrid::smpi::action_test); - xbt_replay_action_register("wait", simgrid::smpi::action_wait); - xbt_replay_action_register("waitAll", simgrid::smpi::action_waitall); - xbt_replay_action_register("barrier", simgrid::smpi::action_barrier); - xbt_replay_action_register("bcast", simgrid::smpi::action_bcast); - xbt_replay_action_register("reduce", simgrid::smpi::action_reduce); - xbt_replay_action_register("allReduce", simgrid::smpi::action_allReduce); - xbt_replay_action_register("allToAll", simgrid::smpi::action_allToAll); - xbt_replay_action_register("allToAllV", simgrid::smpi::action_allToAllv); - xbt_replay_action_register("gather", simgrid::smpi::action_gather); - xbt_replay_action_register("scatter", simgrid::smpi::action_scatter); - xbt_replay_action_register("gatherV", simgrid::smpi::action_gatherv); - xbt_replay_action_register("scatterV", simgrid::smpi::action_scatterv); - xbt_replay_action_register("allGather", simgrid::smpi::action_allgather); - xbt_replay_action_register("allGatherV", simgrid::smpi::action_allgatherv); - xbt_replay_action_register("reduceScatter", simgrid::smpi::action_reducescatter); - xbt_replay_action_register("compute", simgrid::smpi::action_compute); + 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); }); + xbt_replay_action_register("comm_split",[](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::CommunicatorAction().execute(action); }); + xbt_replay_action_register("comm_dup", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::CommunicatorAction().execute(action); }); + xbt_replay_action_register("send", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SendAction("send", storage[simgrid::s4u::this_actor::get_pid()]).execute(action); }); + xbt_replay_action_register("isend", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SendAction("isend", storage[simgrid::s4u::this_actor::get_pid()]).execute(action); }); + 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("bcast", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::BcastAction().execute(action); }); + xbt_replay_action_register("reduce", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ReduceAction().execute(action); }); + xbt_replay_action_register("allreduce", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::AllReduceAction().execute(action); }); + xbt_replay_action_register("alltoall", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::AllToAllAction().execute(action); }); + xbt_replay_action_register("alltoallv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::AllToAllVAction().execute(action); }); + xbt_replay_action_register("gather", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::GatherAction("gather").execute(action); }); + xbt_replay_action_register("scatter", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ScatterAction().execute(action); }); + xbt_replay_action_register("gatherv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::GatherVAction("gatherv").execute(action); }); + xbt_replay_action_register("scatterv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ScatterVAction().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); }); //if we have a delayed start, sleep here. - if(*argc>2){ - double value = xbt_str_parse_double((*argv)[2], "%s is not a double"); - XBT_VERB("Delayed start for instance - Sleeping for %f flops ",value ); - smpi_execute_flops(value); + if (start_delay_flops > 0) { + XBT_VERB("Delayed start for instance - Sleeping for %f flops ", start_delay_flops); + private_execute_flops(start_delay_flops); } else { - //UGLY: force a context switch to be sure that all MSG_processes begin initialization - XBT_DEBUG("Force context switch by smpi_execute_flops - Sleeping for 0.0 flops "); - smpi_execute_flops(0.0); + // 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 */ -void smpi_replay_main(int* argc, char*** argv) +void smpi_replay_main(int rank, const char* private_trace_filename) { - simgrid::xbt::replay_runner(*argc, *argv); + 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(), private_trace_filename); /* and now, finalize everything */ /* One active process will stop. Decrease the counter*/ - XBT_DEBUG("There are %zu elements in reqq[*]", get_reqq_self()->size()); - if (not get_reqq_self()->empty()) { - unsigned int count_requests=get_reqq_self()->size(); - MPI_Request requests[count_requests]; - MPI_Status status[count_requests]; + unsigned int count_requests = storage[simgrid::s4u::this_actor::get_pid()].size(); + 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& req : *get_reqq_self()) { - requests[i] = req; + 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, status); + simgrid::smpi::Request::waitall(count_requests, requests.data(), MPI_STATUSES_IGNORE); } - delete get_reqq_self(); + + if (simgrid::config::get_value("smpi/barrier-finalization")) + simgrid::smpi::colls::barrier(MPI_COMM_WORLD); + active_processes--; if(active_processes==0){ @@ -899,17 +944,17 @@ void smpi_replay_main(int* argc, char*** argv) smpi_free_replay_tmp_buffers(); } - TRACE_smpi_comm_in(Actor::self()->getPid(), "smpi_replay_run_finalize", new simgrid::instr::NoOpTIData("finalize")); + TRACE_smpi_comm_in(simgrid::s4u::this_actor::get_pid(), "smpi_replay_run_finalize", + new simgrid::instr::NoOpTIData("finalize")); smpi_process()->finalize(); - TRACE_smpi_comm_out(Actor::self()->getPid()); - TRACE_smpi_finalize(Actor::self()->getPid()); + TRACE_smpi_comm_out(simgrid::s4u::this_actor::get_pid()); } /** @brief chain a replay initialization and a replay start */ -void smpi_replay_run(int* argc, char*** argv) +void smpi_replay_run(const char* instance_id, int rank, double start_delay_flops, const char* private_trace_filename) { - smpi_replay_init(argc, argv); - smpi_replay_main(argc, argv); + smpi_replay_init(instance_id, rank, start_delay_flops); + smpi_replay_main(rank, private_trace_filename); }