X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/2807fde4fd1f59c230d69a934634c5dfb77905f2..ad1c477db1d600a17fa9f9b95dccefd0ba479ef3:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index 8e02386a1f..33c954c811 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -1,4 +1,4 @@ -/* Copyright (c) 2009-2017. The SimGrid Team. All rights reserved. */ +/* Copyright (c) 2009-2018. 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. */ @@ -12,910 +12,936 @@ #include "smpi_request.hpp" #include "xbt/replay.hpp" +#include +#include +#include #include +#include #include -#define KEY_SIZE (sizeof(int) * 2 + 1) - -XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI"); +using simgrid::s4u::Actor; + +#include +// 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 + struct hash + { + size_t + operator()(TT const& tt) const + { + return std::hash()(tt); + } + }; + + template + inline void hash_combine(std::size_t& seed, T const& v) + { + seed ^= hash_tuple::hash()(v) + 0x9e3779b9 + (seed<<6) + (seed>>2); + } -int communicator_size = 0; -static int active_processes = 0; -std::unordered_map*> reqq; + // Recursive template code derived from Matthieu M. + template ::value - 1> + struct HashValueImpl + { + static void apply(size_t& seed, Tuple const& tuple) + { + HashValueImpl::apply(seed, tuple); + hash_combine(seed, std::get(tuple)); + } + }; + + template + struct HashValueImpl + { + static void apply(size_t& seed, Tuple const& tuple) + { + hash_combine(seed, std::get<0>(tuple)); + } + }; + + template + struct hash> + { + size_t + operator()(std::tuple const& tt) const + { + size_t seed = 0; + HashValueImpl >::apply(seed, tt); + return seed; + } + }; +} -MPI_Datatype MPI_DEFAULT_TYPE; -MPI_Datatype MPI_CURRENT_TYPE; +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI"); -static int sendbuffer_size=0; -char* sendbuffer=nullptr; -static int recvbuffer_size=0; -char* recvbuffer=nullptr; +static std::unordered_map*> reqq; +typedef std::tuple req_key_t; +typedef std::unordered_map>> req_storage_t; + +static MPI_Datatype MPI_DEFAULT_TYPE; + +#define CHECK_ACTION_PARAMS(action, mandatory, optional) \ + { \ + if (action.size() < static_cast(mandatory + 2)) { \ + std::stringstream ss; \ + for (const auto& elem : action) { \ + ss << elem << " "; \ + } \ + THROWF(arg_error, 0, "%s replay failed.\n" \ + "%zu 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" \ + "The full line that was given is:\n %s\n" \ + "Please contact the Simgrid team if support is needed", \ + __func__, action.size(), static_cast(mandatory), static_cast(optional), \ + ss.str().c_str()); \ + } \ + } -static void log_timed_action (const char *const *action, double clock){ +static void log_timed_action(simgrid::xbt::ReplayAction& action, double clock) +{ if (XBT_LOG_ISENABLED(smpi_replay, xbt_log_priority_verbose)){ - char *name = xbt_str_join_array(action, " "); - XBT_VERB("%s %f", name, smpi_process()->simulated_elapsed()-clock); - xbt_free(name); + std::string s = boost::algorithm::join(action, " "); + XBT_VERB("%s %f", s.c_str(), smpi_process()->simulated_elapsed() - clock); } } static std::vector* get_reqq_self() { - return reqq.at(smpi_process()->index()); + return reqq.at(simgrid::s4u::this_actor::get_pid()); } static void set_reqq_self(std::vector *mpi_request) { - reqq.insert({smpi_process()->index(), mpi_request}); -} - -//allocate a single buffer for all sends, growing it if needed -void* smpi_get_tmp_sendbuffer(int size) -{ - if (not smpi_process()->replaying()) - return xbt_malloc(size); - if (sendbuffer_size(xbt_realloc(sendbuffer,size)); - sendbuffer_size=size; - } - return sendbuffer; -} - -//allocate a single buffer for all recv -void* smpi_get_tmp_recvbuffer(int size){ - if (not smpi_process()->replaying()) - return xbt_malloc(size); - if (recvbuffer_size(xbt_realloc(recvbuffer,size)); - recvbuffer_size=size; - } - return recvbuffer; -} - -void smpi_free_tmp_buffer(void* buf){ - if (not smpi_process()->replaying()) - xbt_free(buf); + reqq.insert({simgrid::s4u::this_actor::get_pid(), mpi_request}); } /* Helper function */ -static double parse_double(const char *string) -{ - char *endptr; - double value = strtod(string, &endptr); - if (*endptr != '\0') - THROWF(unknown_error, 0, "%s is not a double", string); - return value; -} - -static MPI_Datatype decode_datatype(const char *const action) -{ - switch(atoi(action)) { - case 0: - MPI_CURRENT_TYPE=MPI_DOUBLE; - break; - case 1: - MPI_CURRENT_TYPE=MPI_INT; - break; - case 2: - MPI_CURRENT_TYPE=MPI_CHAR; - break; - case 3: - MPI_CURRENT_TYPE=MPI_SHORT; - break; - case 4: - MPI_CURRENT_TYPE=MPI_LONG; - break; - case 5: - MPI_CURRENT_TYPE=MPI_FLOAT; - break; - case 6: - MPI_CURRENT_TYPE=MPI_BYTE; - break; - default: - MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; - break; - } - return MPI_CURRENT_TYPE; -} - -const char* encode_datatype(MPI_Datatype datatype) +static double parse_double(std::string string) { - if (datatype==MPI_BYTE) - return ""; - if(datatype==MPI_DOUBLE) - return "0"; - if(datatype==MPI_INT) - return "1"; - if(datatype==MPI_CHAR) - return "2"; - if(datatype==MPI_SHORT) - return "3"; - if(datatype==MPI_LONG) - return "4"; - if(datatype==MPI_FLOAT) - return "5"; - // default - not implemented. - // do not warn here as we pass in this function even for other trace formats - return "-1"; + return xbt_str_parse_double(string.c_str(), "%s is not a double"); } -#define CHECK_ACTION_PARAMS(action, mandatory, optional) {\ - int i=0;\ - while(action[i]!=nullptr)\ - i++;\ - if(isimulated_start(); - /*initialize the number of active processes */ - active_processes = smpi_process_count(); - - set_reqq_self(new std::vector); -} - -static void action_finalize(const char *const *action) -{ - /* Nothing to do */ -} - -static void action_comm_size(const char *const *action) -{ - communicator_size = parse_double(action[2]); - log_timed_action (action, smpi_process()->simulated_elapsed()); -} - -static void action_comm_split(const char *const *action) -{ - log_timed_action (action, smpi_process()->simulated_elapsed()); -} - -static void action_comm_dup(const char *const *action) -{ - log_timed_action (action, smpi_process()->simulated_elapsed()); -} - -static void action_compute(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 1, 0) - double clock = smpi_process()->simulated_elapsed(); - double flops= parse_double(action[2]); - int rank = smpi_process()->index(); - - TRACE_smpi_computing_in(rank, flops); - smpi_execute_flops(flops); - TRACE_smpi_computing_out(rank); - - log_timed_action (action, clock); -} - -static void action_send(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 2, 1) - int to = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - - if(action[4]) - MPI_CURRENT_TYPE=decode_datatype(action[4]); - else - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; - - int rank = smpi_process()->index(); - int dst_traced = MPI_COMM_WORLD->group()->rank(to); - - TRACE_smpi_comm_in(rank, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("send", dst_traced, size, encode_datatype(MPI_CURRENT_TYPE))); - if (not TRACE_smpi_view_internals()) - TRACE_smpi_send(rank, rank, dst_traced, 0, size*MPI_CURRENT_TYPE->size()); +namespace replay { - Request::send(nullptr, size, MPI_CURRENT_TYPE, to , 0, MPI_COMM_WORLD); +class RequestStorage { +private: + req_storage_t store; - TRACE_smpi_comm_out(rank); - - log_timed_action(action, clock); -} - -static void action_Isend(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 2, 1) - int to = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - - if(action[4]) - MPI_CURRENT_TYPE=decode_datatype(action[4]); - else - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; - - int rank = smpi_process()->index(); - int dst_traced = MPI_COMM_WORLD->group()->rank(to); - TRACE_smpi_comm_in(rank, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("Isend", dst_traced, size, encode_datatype(MPI_CURRENT_TYPE))); - if (not TRACE_smpi_view_internals()) - TRACE_smpi_send(rank, rank, 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(rank); +public: + RequestStorage() {} + int size() + { + return store.size(); + } - get_reqq_self()->push_back(request); + req_storage_t& get_store() + { + return store; + } - log_timed_action (action, clock); -} + void get_requests(std::vector& vec) + { + for (auto& pair : store) { + auto& req = pair.second; + auto my_proc_id = simgrid::s4u::this_actor::get_pid(); + if (req != MPI_REQUEST_NULL && (req->src() == my_proc_id || req->dst() == my_proc_id)) { + vec.push_back(pair.second); + pair.second->print_request("MM"); + } + } + } -static void action_recv(const char *const *action) { - CHECK_ACTION_PARAMS(action, 2, 1) - int from = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); - MPI_Status status; + MPI_Request find(int src, int dst, int tag) + { + req_storage_t::iterator it = store.find(req_key_t(src, dst, tag)); + return (it == store.end()) ? MPI_REQUEST_NULL : it->second; + } - if(action[4]) - MPI_CURRENT_TYPE=decode_datatype(action[4]); - else - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; + void remove(MPI_Request req) + { + if (req == MPI_REQUEST_NULL) return; - int rank = smpi_process()->index(); - int src_traced = MPI_COMM_WORLD->group()->rank(from); + store.erase(req_key_t(req->src()-1, req->dst()-1, req->tag())); + } - TRACE_smpi_comm_in(rank, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("recv", src_traced, size, encode_datatype(MPI_CURRENT_TYPE))); + 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}); + } - //unknown size from the receiver point of view - if (size <= 0.0) { - Request::probe(from, 0, MPI_COMM_WORLD, &status); - size=status.count; + /* 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(src, dst, tag), MPI_REQUEST_NULL}); + } +}; + +class ActionArgParser { +public: + virtual ~ActionArgParser() = default; + virtual void parse(simgrid::xbt::ReplayAction& action, std::string name) { CHECK_ACTION_PARAMS(action, 0, 0) } +}; + +class WaitTestParser : public ActionArgParser { +public: + int src; + int dst; + int tag; + + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + CHECK_ACTION_PARAMS(action, 3, 0) + src = std::stoi(action[2]); + dst = std::stoi(action[3]); + tag = std::stoi(action[4]); } - - Request::recv(nullptr, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD, &status); - - TRACE_smpi_comm_out(rank); - if (not TRACE_smpi_view_internals()) { - TRACE_smpi_recv(src_traced, rank, 0); +}; + +class SendRecvParser : public ActionArgParser { +public: + /* communication partner; if we send, this is the receiver and vice versa */ + int partner; + double size; + int tag; + MPI_Datatype datatype1 = MPI_DEFAULT_TYPE; + + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + 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]); } +}; - log_timed_action (action, clock); -} - -static void action_Irecv(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 2, 1) - int from = atoi(action[2]); - double size=parse_double(action[3]); - double clock = smpi_process()->simulated_elapsed(); +class ComputeParser : public ActionArgParser { +public: + /* communication partner; if we send, this is the receiver and vice versa */ + double flops; - if(action[4]) - MPI_CURRENT_TYPE=decode_datatype(action[4]); - else - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; - - int rank = smpi_process()->index(); - int src_traced = MPI_COMM_WORLD->group()->rank(from); - TRACE_smpi_comm_in(rank, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("Irecv", src_traced, size, encode_datatype(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; + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + CHECK_ACTION_PARAMS(action, 1, 0) + flops = parse_double(action[2]); } - - MPI_Request request = Request::irecv(nullptr, size, MPI_CURRENT_TYPE, from, 0, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(rank); - get_reqq_self()->push_back(request); - - log_timed_action (action, clock); -} - -static void action_test(const char* const* action) -{ - 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 rank = smpi_process()->index(); - TRACE_smpi_testing_in(rank); - - 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); - - TRACE_smpi_testing_out(rank); +}; + +class CollCommParser : public ActionArgParser { +public: + double size; + double comm_size; + double comp_size; + int send_size; + int recv_size; + int root = 0; + MPI_Datatype datatype1 = MPI_DEFAULT_TYPE; + MPI_Datatype datatype2 = MPI_DEFAULT_TYPE; +}; + +class BcastArgParser : public CollCommParser { +public: + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + 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]); } - log_timed_action (action, clock); -} - -static void action_wait(const char *const *action){ - CHECK_ACTION_PARAMS(action, 0, 0) - double clock = smpi_process()->simulated_elapsed(); - MPI_Status status; - - xbt_assert(get_reqq_self()->size(), "action wait not preceded by any irecv or isend: %s", - xbt_str_join_array(action," ")); - 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; +}; + +class ReduceArgParser : public CollCommParser { +public: + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + CHECK_ACTION_PARAMS(action, 2, 2) + comm_size = parse_double(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]); } - - 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()); - int is_wait_for_receive = (request->flags() & RECV); - 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); - log_timed_action (action, clock); -} - -static void action_waitall(const char *const *action){ - 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 rank_traced = smpi_process()->index(); - TRACE_smpi_comm_in(rank_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(rank_traced); - } - log_timed_action (action, clock); -} - -static void action_barrier(const char *const *action){ - double clock = smpi_process()->simulated_elapsed(); - int rank = smpi_process()->index(); - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::NoOpTIData("barrier")); - - Colls::barrier(MPI_COMM_WORLD); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} - -static void action_bcast(const char *const *action) -{ - CHECK_ACTION_PARAMS(action, 1, 2) - double size = parse_double(action[2]); - double clock = smpi_process()->simulated_elapsed(); - int root=0; - /* Initialize MPI_CURRENT_TYPE in order to decrease the number of the checks */ - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; - - if(action[3]) { - root= atoi(action[3]); - if(action[4]) - MPI_CURRENT_TYPE=decode_datatype(action[4]); +}; + +class AllReduceArgParser : public CollCommParser { +public: + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + CHECK_ACTION_PARAMS(action, 2, 1) + comm_size = parse_double(action[2]); + comp_size = parse_double(action[3]); + if (action.size() > 4) + datatype1 = simgrid::smpi::Datatype::decode(action[4]); } - - int rank = smpi_process()->index(); - TRACE_smpi_comm_in(rank, __FUNCTION__, - new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->index(root), -1.0, size, -1, - encode_datatype(MPI_CURRENT_TYPE), "")); - - void *sendbuf = smpi_get_tmp_sendbuffer(size* MPI_CURRENT_TYPE->size()); - - Colls::bcast(sendbuf, size, MPI_CURRENT_TYPE, root, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} - -static void action_reduce(const char *const *action) -{ - 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=0; - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; - - if(action[4]) { - root= atoi(action[4]); - if(action[5]) - MPI_CURRENT_TYPE=decode_datatype(action[5]); - } - - int rank = smpi_process()->index(); - TRACE_smpi_comm_in(rank, __FUNCTION__, - new simgrid::instr::CollTIData("reduce", MPI_COMM_WORLD->group()->index(root), comp_size, - comm_size, -1, encode_datatype(MPI_CURRENT_TYPE), "")); - - 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); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} - -static void action_allReduce(const char *const *action) { - CHECK_ACTION_PARAMS(action, 2, 1) - double comm_size = parse_double(action[2]); - double comp_size = parse_double(action[3]); - - if(action[4]) - MPI_CURRENT_TYPE=decode_datatype(action[4]); - else - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; - - double clock = smpi_process()->simulated_elapsed(); - int rank = smpi_process()->index(); - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("allReduce", -1, comp_size, comm_size, -1, - encode_datatype(MPI_CURRENT_TYPE), "")); - - 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::allreduce(sendbuf, recvbuf, comm_size, MPI_CURRENT_TYPE, MPI_OP_NULL, MPI_COMM_WORLD); - smpi_execute_flops(comp_size); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} - -static void action_allToAll(const char *const *action) { - CHECK_ACTION_PARAMS(action, 2, 2) //two mandatory (send and recv volumes) and two optional (corresponding datatypes) - double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); - int send_size = parse_double(action[2]); - int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; - - if(action[4] && action[5]) { - MPI_CURRENT_TYPE=decode_datatype(action[4]); - MPI_CURRENT_TYPE2=decode_datatype(action[5]); +}; + +class AllToAllArgParser : public CollCommParser { +public: + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + 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]); } - else - MPI_CURRENT_TYPE=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 rank = smpi_process()->index(); - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("allToAll", -1, -1.0, send_size, recv_size, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); - - Colls::alltoall(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} +}; + +class GatherArgParser : public CollCommParser { +public: + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + /* 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_double(action[2]); + recv_size = parse_double(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]); + } + else { + if (action.size() > 4) + datatype1 = simgrid::smpi::Datatype::decode(action[4]); + if (action.size() > 5) + datatype2 = simgrid::smpi::Datatype::decode(action[5]); + } + } +}; + +class GatherVArgParser : public CollCommParser { +public: + int recv_size_sum; + std::shared_ptr> recvcounts; + std::vector disps; + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + /* 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_double(action[2]); + disps = std::vector(comm_size, 0); + recvcounts = std::shared_ptr>(new std::vector(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]); + } + else { + int datatype_index = 0; + int 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 */ + 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 */ + disp_index = 3 + comm_size; + } else if (action.size() > 3 + comm_size) { /* only datatype, no disp specified */ + datatype_index = 3 + comm_size; + datatype1 = simgrid::smpi::Datatype::decode(action[datatype_index]); + datatype2 = simgrid::smpi::Datatype::decode(action[datatype_index]); + } + + if (disp_index != 0) { + for (unsigned int i = 0; i < comm_size; i++) + disps[i] = std::stoi(action[disp_index + i]); + } + } -static void action_gather(const char *const *action) { - /* 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 + 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); + } +}; + +class ScatterArgParser : public CollCommParser { +public: + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + /* 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) + 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]); + } +}; + +class ScatterVArgParser : public CollCommParser { +public: + int recv_size_sum; + int send_size_sum; + std::shared_ptr> sendcounts; + std::vector disps; + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + /* 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 decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() - */ - CHECK_ACTION_PARAMS(action, 2, 3) - double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); - int send_size = parse_double(action[2]); - int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; - if(action[4] && action[5]) { - MPI_CURRENT_TYPE=decode_datatype(action[5]); - MPI_CURRENT_TYPE2=decode_datatype(action[6]); - } else { - MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; + 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, comm_size + 1, 2) + recv_size = parse_double(action[2 + comm_size]); + disps = std::vector(comm_size, 0); + sendcounts = std::shared_ptr>(new std::vector(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]); + + 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; + } +}; + +class ReduceScatterArgParser : public CollCommParser { +public: + int recv_size_sum; + std::shared_ptr> recvcounts; + std::vector disps; + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + /* 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::shared_ptr>(new std::vector(comm_size)); + if (action.size() > 3 + comm_size) + datatype1 = simgrid::smpi::Datatype::decode(action[3 + comm_size]); + + for (unsigned int i = 0; i < comm_size; i++) { + recvcounts->push_back(std::stoi(action[i + 2])); + } + recv_size_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); + } +}; + +class AllToAllVArgParser : public CollCommParser { +public: + int recv_size_sum; + int send_size_sum; + std::shared_ptr> recvcounts; + std::shared_ptr> sendcounts; + std::vector senddisps; + std::vector recvdisps; + int send_buf_size; + int recv_buf_size; + void parse(simgrid::xbt::ReplayAction& action, std::string name) override + { + /* 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::shared_ptr>(new std::vector(comm_size)); + recvcounts = std::shared_ptr>(new std::vector(comm_size)); + 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]); + + send_buf_size=parse_double(action[2]); + recv_buf_size=parse_double(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]); + } + send_size_sum = std::accumulate(sendcounts->begin(), sendcounts->end(), 0); + recv_size_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0); + } +}; + +template class ReplayAction { +protected: + const std::string name; + RequestStorage* req_storage; // Points to the right storage for this process, nullptr except for Send/Recv/Wait/Test actions. + const int my_proc_id; + T args; + +public: + explicit ReplayAction(std::string name, RequestStorage& storage) : name(name), req_storage(&storage), my_proc_id(simgrid::s4u::this_actor::get_pid()) {} + explicit ReplayAction(std::string name) : name(name), req_storage(nullptr), my_proc_id(simgrid::s4u::this_actor::get_pid()) {} + virtual ~ReplayAction() = default; + + virtual void execute(simgrid::xbt::ReplayAction& action) + { + // Needs to be re-initialized for every action, hence here + double start_time = smpi_process()->simulated_elapsed(); + args.parse(action, name); + kernel(action); + if (name != "Init") + log_timed_action(action, start_time); } - void *send = smpi_get_tmp_sendbuffer(send_size* MPI_CURRENT_TYPE->size()); - void *recv = nullptr; - int root=0; - if(action[4]) - root=atoi(action[4]); - 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, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); - Colls::gather(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); + virtual void kernel(simgrid::xbt::ReplayAction& action) = 0; - TRACE_smpi_comm_out(smpi_process()->index()); - log_timed_action (action, clock); -} + void* send_buffer(int size) + { + return smpi_get_tmp_sendbuffer(size); + } -static void action_scatter(const char* const* action) -{ - /* 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 decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() - */ - CHECK_ACTION_PARAMS(action, 2, 3) - double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); - int send_size = parse_double(action[2]); - int recv_size = parse_double(action[3]); - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; - if (action[4] && action[5]) { - MPI_CURRENT_TYPE = decode_datatype(action[5]); - MPI_CURRENT_TYPE2 = decode_datatype(action[6]); - } else { - MPI_CURRENT_TYPE = MPI_DEFAULT_TYPE; + void* recv_buffer(int size) + { + return smpi_get_tmp_recvbuffer(size); } - void* send = smpi_get_tmp_sendbuffer(send_size * MPI_CURRENT_TYPE->size()); - void* recv = nullptr; - int root = 0; - if (action[4]) - root = atoi(action[4]); - int rank = MPI_COMM_WORLD->rank(); +}; + +class WaitAction : public ReplayAction { +public: + WaitAction(RequestStorage& storage) : ReplayAction("Wait", storage) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + 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()); + MPI_Request request = req_storage->find(args.src, args.dst, args.tag); + req_storage->remove(request); + + if (request == MPI_REQUEST_NULL) { + /* Assume that the trace is well formed, meaning the comm might have been caught by a MPI_test. Then just + * return.*/ + return; + } - if (rank == root) - recv = smpi_get_tmp_recvbuffer(recv_size * comm_size * MPI_CURRENT_TYPE2->size()); + int rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1; - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("gather", root, -1.0, send_size, recv_size, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + // 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() & RECV); + // TODO: Here we take the rank while we normally take the process id (look for my_proc_id) + TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::NoOpTIData("wait")); - Colls::scatter(send, send_size, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); + MPI_Status status; + Request::wait(&request, &status); - TRACE_smpi_comm_out(smpi_process()->index()); - log_timed_action(action, clock); -} + TRACE_smpi_comm_out(rank); + if (is_wait_for_receive) + TRACE_smpi_recv(args.src, args.dst, args.tag); + } +}; + +class SendAction : public ReplayAction { +public: + SendAction() = delete; + explicit SendAction(std::string name, RequestStorage& storage) : ReplayAction(name, storage) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); + + TRACE_smpi_comm_in(my_proc_id, __func__, new simgrid::instr::Pt2PtTIData(name, args.partner, args.size, + args.tag, Datatype::encode(args.datatype1))); + if (not TRACE_smpi_view_internals()) + TRACE_smpi_send(my_proc_id, my_proc_id, dst_traced, args.tag, args.size * args.datatype1->size()); + + if (name == "send") { + Request::send(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD); + } else if (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", name.c_str()); + } -static void action_gatherv(const char *const *action) { - /* 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 decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() - */ - double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size+1, 2) - int send_size = parse_double(action[2]); - int disps[comm_size]; - int recvcounts[comm_size]; - int recv_sum=0; - - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; - if(action[4+comm_size] && action[5+comm_size]) { - MPI_CURRENT_TYPE=decode_datatype(action[4+comm_size]); - MPI_CURRENT_TYPE2=decode_datatype(action[5+comm_size]); - } else - MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; - - void *send = smpi_get_tmp_sendbuffer(send_size* MPI_CURRENT_TYPE->size()); - void *recv = nullptr; - for(int i=0;irank(); - - if(rank==root) - recv = smpi_get_tmp_recvbuffer(recv_sum* MPI_CURRENT_TYPE2->size()); - - std::vector* trace_recvcounts = new std::vector; - for (int i = 0; i < comm_size; i++) // copy data to avoid bad free - trace_recvcounts->push_back(recvcounts[i]); - - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData( - "gatherV", root, send_size, nullptr, -1, trace_recvcounts, - encode_datatype(MPI_CURRENT_TYPE), encode_datatype(MPI_CURRENT_TYPE2))); - - Colls::gatherv(send, send_size, MPI_CURRENT_TYPE, recv, recvcounts, disps, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(smpi_process()->index()); - log_timed_action (action, clock); -} + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class RecvAction : public ReplayAction { +public: + RecvAction() = delete; + explicit RecvAction(std::string name, RequestStorage& storage) : ReplayAction(name, storage) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + int src_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); + + TRACE_smpi_comm_in(my_proc_id, __func__, new simgrid::instr::Pt2PtTIData(name, args.partner, args.size, + args.tag, Datatype::encode(args.datatype1))); + + MPI_Status status; + // unknown size from the receiver point of view + if (args.size <= 0.0) { + Request::probe(args.partner, args.tag, MPI_COMM_WORLD, &status); + args.size = status.count; + } -static void action_scatterv(const char* const* action) -{ - /* 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 decode_datatype() - 5) 0 is the recv datatype id, see decode_datatype() - */ - double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size + 1, 2) - int recv_size = parse_double(action[2 + comm_size]); - int disps[comm_size]; - int sendcounts[comm_size]; - int send_sum = 0; - - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; - if (action[4 + comm_size] && action[5 + comm_size]) { - MPI_CURRENT_TYPE = decode_datatype(action[4 + comm_size]); - MPI_CURRENT_TYPE2 = decode_datatype(action[5 + comm_size]); - } else - MPI_CURRENT_TYPE = MPI_DEFAULT_TYPE; - - void* send = nullptr; - void* recv = smpi_get_tmp_recvbuffer(recv_size * MPI_CURRENT_TYPE->size()); - for (int i = 0; i < comm_size; i++) { - sendcounts[i] = atoi(action[i + 2]); - send_sum += sendcounts[i]; - disps[i] = 0; - } - - int root = atoi(action[3 + comm_size]); - int rank = MPI_COMM_WORLD->rank(); - - if (rank == root) - send = smpi_get_tmp_sendbuffer(send_sum * MPI_CURRENT_TYPE2->size()); - - std::vector* trace_sendcounts = new std::vector; - for (int i = 0; i < comm_size; i++) // copy data to avoid bad free - trace_sendcounts->push_back(sendcounts[i]); - - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData( - "gatherV", root, -1, trace_sendcounts, recv_size, nullptr, - encode_datatype(MPI_CURRENT_TYPE), encode_datatype(MPI_CURRENT_TYPE2))); - - Colls::scatterv(send, sendcounts, disps, MPI_CURRENT_TYPE, recv, recv_size, MPI_CURRENT_TYPE2, root, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(smpi_process()->index()); - log_timed_action(action, clock); -} + if (name == "recv") { + Request::recv(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD, &status); + } else if (name == "Irecv") { + MPI_Request request = Request::irecv(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD); + req_storage->add(request); + } -static void action_reducescatter(const char *const *action) { - /* 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 decode_datatype(). -*/ - double clock = smpi_process()->simulated_elapsed(); - int comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size+1, 1) - int comp_size = parse_double(action[2+comm_size]); - int recvcounts[comm_size]; - int rank = smpi_process()->index(); - int size = 0; - std::vector* trace_recvcounts = new std::vector; - if(action[3+comm_size]) - MPI_CURRENT_TYPE=decode_datatype(action[3+comm_size]); - else - MPI_CURRENT_TYPE= MPI_DEFAULT_TYPE; - - for(int i=0;ipush_back(recvcounts[i]); - size+=recvcounts[i]; - } - - TRACE_smpi_comm_in(rank, __FUNCTION__, - new simgrid::instr::VarCollTIData("reduceScatter", -1, 0, nullptr, -1, trace_recvcounts, - std::to_string(comp_size), /* ugly hack to print comp_size */ - encode_datatype(MPI_CURRENT_TYPE))); - - void *sendbuf = smpi_get_tmp_sendbuffer(size* MPI_CURRENT_TYPE->size()); - void *recvbuf = smpi_get_tmp_recvbuffer(size* MPI_CURRENT_TYPE->size()); - - Colls::reduce_scatter(sendbuf, recvbuf, recvcounts, MPI_CURRENT_TYPE, MPI_OP_NULL, MPI_COMM_WORLD); - smpi_execute_flops(comp_size); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} + TRACE_smpi_comm_out(my_proc_id); + // TODO: Check why this was only activated in the "recv" case and not in the "Irecv" case + if (name == "recv" && not TRACE_smpi_view_internals()) { + TRACE_smpi_recv(src_traced, my_proc_id, args.tag); + } + } +}; + +class ComputeAction : public ReplayAction { +public: + ComputeAction() : ReplayAction("compute") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_computing_in(my_proc_id, args.flops); + smpi_execute_flops(args.flops); + TRACE_smpi_computing_out(my_proc_id); + } +}; + +class TestAction : public ReplayAction { +public: + TestAction(RequestStorage& storage) : ReplayAction("Test", storage) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + MPI_Request request = req_storage->find(args.src, args.dst, args.tag); + req_storage->remove(request); + // 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_testing_in(my_proc_id); + + MPI_Status status; + 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.*/ + if (request == MPI_REQUEST_NULL) + req_storage->addNullRequest(args.src, args.dst, args.tag); + else + req_storage->add(request); + + TRACE_smpi_testing_out(my_proc_id); + } + } +}; + +class InitAction : public ReplayAction { +public: + InitAction() : ReplayAction("Init") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + CHECK_ACTION_PARAMS(action, 0, 1) + MPI_DEFAULT_TYPE = (action.size() > 2) ? MPI_DOUBLE // default MPE datatype + : MPI_BYTE; // default TAU datatype + + /* start a simulated timer */ + smpi_process()->simulated_start(); + } +}; + +class CommunicatorAction : public ReplayAction { +public: + CommunicatorAction() : ReplayAction("Comm") {} + void kernel(simgrid::xbt::ReplayAction& action) override { /* nothing to do */} +}; + +class WaitAllAction : public ReplayAction { +public: + WaitAllAction(RequestStorage& storage) : ReplayAction("waitAll", storage) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + const unsigned int count_requests = req_storage->size(); + + if (count_requests > 0) { + TRACE_smpi_comm_in(my_proc_id, __func__, new simgrid::instr::Pt2PtTIData("waitAll", -1, count_requests, "")); + std::vector> sender_receiver; + std::vector reqs; + req_storage->get_requests(reqs); + for (const auto& req : reqs) { + if (req && (req->flags() & RECV)) { + sender_receiver.push_back({req->src(), req->dst()}); + } + } + MPI_Status status[count_requests]; + Request::waitall(count_requests, &(reqs.data())[0], status); + + for (auto& pair : sender_receiver) { + TRACE_smpi_recv(pair.first, pair.second, 0); + } + TRACE_smpi_comm_out(my_proc_id); + } + } +}; + +class BarrierAction : public ReplayAction { +public: + BarrierAction() : ReplayAction("barrier") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, __func__, new simgrid::instr::NoOpTIData("barrier")); + Colls::barrier(MPI_COMM_WORLD); + TRACE_smpi_comm_out(my_proc_id); + } +}; -static void action_allgather(const char *const *action) { - /* 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 decode_datatype(). - */ - double clock = smpi_process()->simulated_elapsed(); +class BcastAction : public ReplayAction { +public: + BcastAction() : ReplayAction("bcast") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, "action_bcast", + new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->actor(args.root)->get_pid(), + -1.0, args.size, -1, Datatype::encode(args.datatype1), "")); - CHECK_ACTION_PARAMS(action, 2, 2) - int sendcount=atoi(action[2]); - int recvcount=atoi(action[3]); + Colls::bcast(send_buffer(args.size * args.datatype1->size()), args.size, args.datatype1, args.root, MPI_COMM_WORLD); - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class ReduceAction : public ReplayAction { +public: + ReduceAction() : ReplayAction("reduce") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, "action_reduce", + new simgrid::instr::CollTIData("reduce", MPI_COMM_WORLD->group()->actor(args.root)->get_pid(), + args.comp_size, args.comm_size, -1, + Datatype::encode(args.datatype1), "")); + + 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); + smpi_execute_flops(args.comp_size); + + TRACE_smpi_comm_out(my_proc_id); + } +}; - if(action[4] && action[5]) { - MPI_CURRENT_TYPE = decode_datatype(action[4]); - MPI_CURRENT_TYPE2 = decode_datatype(action[5]); - } else - MPI_CURRENT_TYPE = MPI_DEFAULT_TYPE; +class AllReduceAction : public ReplayAction { +public: + AllReduceAction() : ReplayAction("allReduce") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, "action_allReduce", new simgrid::instr::CollTIData("allReduce", -1, args.comp_size, args.comm_size, -1, + Datatype::encode(args.datatype1), "")); - void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* MPI_CURRENT_TYPE->size()); - void *recvbuf = smpi_get_tmp_recvbuffer(recvcount* MPI_CURRENT_TYPE2->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); + smpi_execute_flops(args.comp_size); - int rank = smpi_process()->index(); + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class AllToAllAction : public ReplayAction { +public: + AllToAllAction() : ReplayAction("allToAll") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, "action_allToAll", + new simgrid::instr::CollTIData("allToAll", -1, -1.0, args.send_size, args.recv_size, + 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, MPI_COMM_WORLD); + + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class GatherAction : public ReplayAction { +public: + explicit GatherAction(std::string name) : ReplayAction(name) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, name.c_str(), new simgrid::instr::CollTIData(name, (name == "gather") ? args.root : -1, -1.0, args.send_size, args.recv_size, + Datatype::encode(args.datatype1), Datatype::encode(args.datatype2))); + + if (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, 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_in(rank, __FUNCTION__, new simgrid::instr::CollTIData("allGather", -1, -1.0, sendcount, recvcount, - encode_datatype(MPI_CURRENT_TYPE), - encode_datatype(MPI_CURRENT_TYPE2))); + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class GatherVAction : public ReplayAction { +public: + explicit GatherVAction(std::string name) : ReplayAction(name) {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + int rank = MPI_COMM_WORLD->rank(); + + TRACE_smpi_comm_in(my_proc_id, name.c_str(), new simgrid::instr::VarCollTIData( + name, (name == "gatherV") ? args.root : -1, args.send_size, nullptr, -1, args.recvcounts, + Datatype::encode(args.datatype1), Datatype::encode(args.datatype2))); + + if (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); + } - Colls::allgather(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcount, MPI_CURRENT_TYPE2, MPI_COMM_WORLD); + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class ScatterAction : public ReplayAction { +public: + ScatterAction() : ReplayAction("scatter") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + int rank = MPI_COMM_WORLD->rank(); + TRACE_smpi_comm_in(my_proc_id, "action_scatter", new simgrid::instr::CollTIData(name, args.root, -1.0, args.send_size, args.recv_size, + Datatype::encode(args.datatype1), + Datatype::encode(args.datatype2))); + + 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); + + TRACE_smpi_comm_out(my_proc_id); + } +}; - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} -static void action_allgatherv(const char *const *action) { - /* 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 decode_datatype(). - */ - double clock = smpi_process()->simulated_elapsed(); - - int comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, comm_size+1, 2) - int sendcount=atoi(action[2]); - int recvcounts[comm_size]; - int disps[comm_size]; - int recv_sum=0; - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; - - if(action[3+comm_size] && action[4+comm_size]) { - MPI_CURRENT_TYPE = decode_datatype(action[3+comm_size]); - MPI_CURRENT_TYPE2 = decode_datatype(action[4+comm_size]); - } else - MPI_CURRENT_TYPE = MPI_DEFAULT_TYPE; - - void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* MPI_CURRENT_TYPE->size()); - - for(int i=0;isize()); - - int rank = smpi_process()->index(); - - std::vector* trace_recvcounts = new std::vector; - for (int i = 0; i < comm_size; i++) // copy data to avoid bad free - trace_recvcounts->push_back(recvcounts[i]); - - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData( - "allGatherV", -1, sendcount, nullptr, -1, trace_recvcounts, - encode_datatype(MPI_CURRENT_TYPE), encode_datatype(MPI_CURRENT_TYPE2))); - - Colls::allgatherv(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcounts, disps, MPI_CURRENT_TYPE2, - MPI_COMM_WORLD); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} +class ScatterVAction : public ReplayAction { +public: + ScatterVAction() : ReplayAction("scatterV") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + int rank = MPI_COMM_WORLD->rank(); + TRACE_smpi_comm_in(my_proc_id, "action_scatterv", new simgrid::instr::VarCollTIData(name, args.root, -1, args.sendcounts, args.recv_size, + nullptr, Datatype::encode(args.datatype1), + Datatype::encode(args.datatype2))); -static void action_allToAllv(const char *const *action) { - /* 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(); - - int comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, 2*comm_size+2, 2) - int send_size = 0; - int recv_size = 0; - int sendcounts[comm_size]; - std::vector* trace_sendcounts = new std::vector; - int recvcounts[comm_size]; - std::vector* trace_recvcounts = new std::vector; - int senddisps[comm_size]; - int recvdisps[comm_size]; - - MPI_Datatype MPI_CURRENT_TYPE2 = MPI_DEFAULT_TYPE; - - int send_buf_size=parse_double(action[2]); - int recv_buf_size=parse_double(action[3+comm_size]); - if(action[4+2*comm_size] && action[5+2*comm_size]) { - MPI_CURRENT_TYPE=decode_datatype(action[4+2*comm_size]); - MPI_CURRENT_TYPE2=decode_datatype(action[5+2*comm_size]); - } - else - MPI_CURRENT_TYPE=MPI_DEFAULT_TYPE; - - int rank = smpi_process()->index(); - 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(int i=0;ipush_back(sendcounts[i]); - send_size += sendcounts[i]; - recvcounts[i] = atoi(action[i+4+comm_size]); - trace_recvcounts->push_back(recvcounts[i]); - recv_size += recvcounts[i]; - senddisps[i] = 0; - recvdisps[i] = 0; - } - - TRACE_smpi_comm_in(rank, __FUNCTION__, new simgrid::instr::VarCollTIData( - "allToAllV", -1, send_size, trace_sendcounts, recv_size, trace_recvcounts, - encode_datatype(MPI_CURRENT_TYPE), encode_datatype(MPI_CURRENT_TYPE2))); - - Colls::alltoallv(sendbuf, sendcounts, senddisps, MPI_CURRENT_TYPE,recvbuf, recvcounts, recvdisps, - MPI_CURRENT_TYPE, MPI_COMM_WORLD); - - TRACE_smpi_comm_out(rank); - log_timed_action (action, clock); -} + 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); + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class ReduceScatterAction : public ReplayAction { +public: + ReduceScatterAction() : ReplayAction("reduceScatter") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, "action_reducescatter", + new simgrid::instr::VarCollTIData("reduceScatter", -1, 0, nullptr, -1, args.recvcounts, + std::to_string(args.comp_size), /* ugly hack to print 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); + + smpi_execute_flops(args.comp_size); + TRACE_smpi_comm_out(my_proc_id); + } +}; + +class AllToAllVAction : public ReplayAction { +public: + AllToAllVAction() : ReplayAction("allToAllV") {} + void kernel(simgrid::xbt::ReplayAction& action) override + { + TRACE_smpi_comm_in(my_proc_id, __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(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); + } +}; +} // Replay Namespace }} // namespace simgrid::smpi +std::vector storage; /** @brief Only initialize the replay, don't do it for real */ void smpi_replay_init(int* argc, char*** argv) { @@ -923,37 +949,40 @@ void smpi_replay_init(int* argc, char*** argv) smpi_process()->mark_as_initialized(); smpi_process()->set_replaying(true); - int rank = smpi_process()->index(); - TRACE_smpi_init(rank); - TRACE_smpi_computing_init(rank); - TRACE_smpi_comm_in(rank, "smpi_replay_run_init", new simgrid::instr::NoOpTIData("init")); - TRACE_smpi_comm_out(rank); - 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); + int my_proc_id = simgrid::s4u::this_actor::get_pid(); + storage.resize(smpi_process_count()); + + 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::xbt::ReplayAction& action) { simgrid::smpi::replay::InitAction().execute(action); }); + xbt_replay_action_register("finalize", [](simgrid::xbt::ReplayAction& action) { /* 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()-1]).execute(action); }); + xbt_replay_action_register("Isend", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SendAction("Isend", storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); }); + xbt_replay_action_register("recv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::RecvAction("recv", storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); }); + xbt_replay_action_register("Irecv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::RecvAction("Irecv", storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); }); + xbt_replay_action_register("test", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::TestAction(storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); }); + xbt_replay_action_register("wait", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::WaitAction(storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); }); + xbt_replay_action_register("waitAll", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::WaitAllAction(storage[simgrid::s4u::this_actor::get_pid()-1]).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("compute", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::ComputeAction().execute(action); }); //if we have a delayed start, sleep here. if(*argc>2){ @@ -970,39 +999,40 @@ void smpi_replay_init(int* argc, char*** argv) /** @brief actually run the replay after initialization */ void smpi_replay_main(int* argc, char*** argv) { + static int active_processes = 0; + active_processes++; simgrid::xbt::replay_runner(*argc, *argv); /* 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(); + unsigned int count_requests = storage[simgrid::s4u::this_actor::get_pid() - 1].size(); + XBT_DEBUG("There are %ud elements in reqq[*]", count_requests); + if (count_requests > 0) { MPI_Request requests[count_requests]; MPI_Status status[count_requests]; unsigned int i=0; - for (auto const& req : *get_reqq_self()) { - requests[i] = req; + for (auto const& pair : storage[simgrid::s4u::this_actor::get_pid() - 1].get_store()) { + requests[i] = pair.second; i++; } simgrid::smpi::Request::waitall(count_requests, requests, status); } - delete get_reqq_self(); active_processes--; if(active_processes==0){ /* Last process alive speaking: end the simulated timer */ XBT_INFO("Simulation time %f", smpi_process()->simulated_elapsed()); - xbt_free(sendbuffer); - xbt_free(recvbuffer); + smpi_free_replay_tmp_buffers(); } - TRACE_smpi_comm_in(smpi_process()->index(), "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(smpi_process()->index()); - TRACE_smpi_finalize(smpi_process()->index()); + TRACE_smpi_comm_out(simgrid::s4u::this_actor::get_pid()); + TRACE_smpi_finalize(simgrid::s4u::this_actor::get_pid()); } /** @brief chain a replay initialization and a replay start */