X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/9a9d4426f9baa07b062abc3efc7df532ac61baf2..8dc3394c8bcc3849bbda44a668a714596e714a18:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index 8f0caad0f3..3e4ec9b6b1 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. */ @@ -16,26 +16,86 @@ #include #include #include +#include #include 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); + } + + // 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; + } + }; +} + XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI"); -static int active_processes = 0; static std::unordered_map*> reqq; static MPI_Datatype MPI_DEFAULT_TYPE; #define CHECK_ACTION_PARAMS(action, mandatory, optional) \ { \ - if (action.size() < static_cast(mandatory + 2)) \ + 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" \ - "%lu items were given on the line. First two should be process_id and action. " \ + "%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", \ - __FUNCTION__, action.size(), static_cast(mandatory), \ - static_cast(optional)); \ + __func__, action.size(), static_cast(mandatory), static_cast(optional), \ + ss.str().c_str()); \ + } \ } static void log_timed_action(simgrid::xbt::ReplayAction& action, double clock) @@ -48,12 +108,12 @@ static void log_timed_action(simgrid::xbt::ReplayAction& action, double clock) static std::vector* get_reqq_self() { - return reqq.at(Actor::self()->getPid()); + return reqq.at(simgrid::s4u::this_actor::get_pid()); } static void set_reqq_self(std::vector *mpi_request) { - reqq.insert({Actor::self()->getPid(), mpi_request}); + reqq.insert({simgrid::s4u::this_actor::get_pid(), mpi_request}); } /* Helper function */ @@ -65,10 +125,81 @@ static double parse_double(std::string string) namespace simgrid { namespace smpi { -namespace Replay { +namespace replay { + +class RequestStorage { +private: + req_storage_t store; + +public: + RequestStorage() {} + int size() + { + return store.size(); + } + + req_storage_t& get_store() + { + return store; + } + + void get_requests(std::vector& vec) + { + for (auto& pair : store) { + auto& req = pair.second; + auto my_proc_id = simgrid::s4u::this_actor::getPid(); + 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"); + } + } + } + + 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; + } + + void remove(MPI_Request req) + { + if (req == MPI_REQUEST_NULL) return; + + store.erase(req_key_t(req->src()-1, req->dst()-1, req->tag())); + } + + 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}); + } + + /* 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 void parse(simgrid::xbt::ReplayAction& action) { CHECK_ACTION_PARAMS(action, 0, 0) } + 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]); + } }; class SendRecvParser : public ActionArgParser { @@ -76,15 +207,17 @@ 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) override + void parse(simgrid::xbt::ReplayAction& action, std::string name) override { - CHECK_ACTION_PARAMS(action, 2, 1) + CHECK_ACTION_PARAMS(action, 3, 1) partner = std::stoi(action[2]); - size = parse_double(action[3]); - if (action.size() > 4) - datatype1 = simgrid::smpi::Datatype::decode(action[4]); + tag = std::stoi(action[3]); + size = parse_double(action[4]); + if (action.size() > 5) + datatype1 = simgrid::smpi::Datatype::decode(action[5]); } }; @@ -93,7 +226,7 @@ public: /* communication partner; if we send, this is the receiver and vice versa */ double flops; - void parse(simgrid::xbt::ReplayAction& action) override + void parse(simgrid::xbt::ReplayAction& action, std::string name) override { CHECK_ACTION_PARAMS(action, 1, 0) flops = parse_double(action[2]); @@ -114,7 +247,7 @@ public: class BcastArgParser : public CollCommParser { public: - void parse(simgrid::xbt::ReplayAction& action) override + void parse(simgrid::xbt::ReplayAction& action, std::string name) override { CHECK_ACTION_PARAMS(action, 1, 2) size = parse_double(action[2]); @@ -126,7 +259,7 @@ public: class ReduceArgParser : public CollCommParser { public: - void parse(simgrid::xbt::ReplayAction& action) override + void parse(simgrid::xbt::ReplayAction& action, std::string name) override { CHECK_ACTION_PARAMS(action, 2, 2) comm_size = parse_double(action[2]); @@ -139,7 +272,7 @@ public: class AllReduceArgParser : public CollCommParser { public: - void parse(simgrid::xbt::ReplayAction& action) override + void parse(simgrid::xbt::ReplayAction& action, std::string name) override { CHECK_ACTION_PARAMS(action, 2, 1) comm_size = parse_double(action[2]); @@ -151,7 +284,7 @@ public: class AllToAllArgParser : public CollCommParser { public: - void parse(simgrid::xbt::ReplayAction& action) override + void parse(simgrid::xbt::ReplayAction& action, std::string name) override { CHECK_ACTION_PARAMS(action, 2, 1) comm_size = MPI_COMM_WORLD->size(); @@ -165,9 +298,9 @@ public: } }; -class GatherParser : public CollCommParser { +class GatherArgParser : public CollCommParser { public: - void parse(simgrid::xbt::ReplayAction& action) override + 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 @@ -182,8 +315,102 @@ public: 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 (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]); + } + } + + 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) @@ -191,21 +418,127 @@ public: } }; +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 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; + const int my_proc_id; T args; - int my_proc_id; - public: - explicit ReplayAction(std::string name) : name(name), my_proc_id(simgrid::s4u::Actor::self()->getPid()) {} + explicit ReplayAction(std::string name) : name(name), 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); + args.parse(action, name); kernel(action); if (name != "Init") log_timed_action(action, start_time); @@ -224,7 +557,7 @@ public: } }; -class WaitAction : public ReplayAction { +class WaitAction : public ReplayAction { public: WaitAction() : ReplayAction("Wait") {} void kernel(simgrid::xbt::ReplayAction& action) override @@ -234,7 +567,7 @@ public: MPI_Request request = get_reqq_self()->back(); get_reqq_self()->pop_back(); - if (request == nullptr) { + 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; @@ -244,38 +577,36 @@ public: // Must be taken before Request::wait() since the request may be set to // MPI_REQUEST_NULL by Request::wait! - int src = request->comm()->group()->rank(request->src()); - int dst = request->comm()->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")); + TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::NoOpTIData("wait")); MPI_Status status; Request::wait(&request, &status); TRACE_smpi_comm_out(rank); if (is_wait_for_receive) - TRACE_smpi_recv(src, dst, 0); + TRACE_smpi_recv(args.src, args.dst, args.tag); } }; class SendAction : public ReplayAction { public: SendAction() = delete; - SendAction(std::string name) : ReplayAction(name) {} + explicit SendAction(std::string name) : ReplayAction(name) {} void kernel(simgrid::xbt::ReplayAction& action) override { - int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner)->getPid(); + int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::Pt2PtTIData(name, args.partner, args.size, - Datatype::encode(args.datatype1))); + 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, 0, args.size * args.datatype1->size()); + 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, 0, MPI_COMM_WORLD); + 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, 0, MPI_COMM_WORLD); + MPI_Request request = Request::isend(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD); get_reqq_self()->push_back(request); } else { xbt_die("Don't know this action, %s", name.c_str()); @@ -291,29 +622,29 @@ public: explicit RecvAction(std::string name) : ReplayAction(name) {} void kernel(simgrid::xbt::ReplayAction& action) override { - int src_traced = MPI_COMM_WORLD->group()->actor(args.partner)->getPid(); + int src_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::Pt2PtTIData(name, args.partner, args.size, - Datatype::encode(args.datatype1))); + 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, 0, MPI_COMM_WORLD, &status); + Request::probe(args.partner, args.tag, MPI_COMM_WORLD, &status); args.size = status.count; } if (name == "recv") { - Request::recv(nullptr, args.size, args.datatype1, args.partner, 0, MPI_COMM_WORLD, &status); + 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, 0, MPI_COMM_WORLD); + MPI_Request request = Request::irecv(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD); get_reqq_self()->push_back(request); } 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, 0); + TRACE_smpi_recv(src_traced, my_proc_id, args.tag); } } }; @@ -329,7 +660,7 @@ public: } }; -class TestAction : public ReplayAction { +class TestAction : public ReplayAction { public: TestAction() : ReplayAction("Test") {} void kernel(simgrid::xbt::ReplayAction& action) override @@ -366,9 +697,6 @@ public: /* 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); } }; @@ -387,8 +715,7 @@ public: const unsigned int count_requests = get_reqq_self()->size(); if (count_requests > 0) { - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, - new simgrid::instr::Pt2PtTIData("waitAll", -1, count_requests, "")); + TRACE_smpi_comm_in(my_proc_id, __func__, new simgrid::instr::Pt2PtTIData("waitAll", -1, count_requests, "")); std::vector> sender_receiver; for (const auto& req : (*get_reqq_self())) { if (req && (req->flags() & RECV)) { @@ -411,7 +738,7 @@ public: BarrierAction() : ReplayAction("barrier") {} void kernel(simgrid::xbt::ReplayAction& action) override { - TRACE_smpi_comm_in(my_proc_id, __FUNCTION__, new simgrid::instr::NoOpTIData("barrier")); + 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); } @@ -423,7 +750,7 @@ public: 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)->getPid(), + new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->actor(args.root)->get_pid(), -1.0, args.size, -1, Datatype::encode(args.datatype1), "")); Colls::bcast(send_buffer(args.size * args.datatype1->size()), args.size, args.datatype1, args.root, MPI_COMM_WORLD); @@ -438,8 +765,9 @@ public: 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)->getPid(), args.comp_size, - args.comm_size, -1, Datatype::encode(args.datatype1), "")); + 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); @@ -475,9 +803,9 @@ public: 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); + 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); } @@ -485,15 +813,17 @@ public: class GatherAction : public ReplayAction { public: - GatherAction(std::string name) : ReplayAction(name) {} + explicit GatherAction(std::string name) : ReplayAction(name) {} void kernel(simgrid::xbt::ReplayAction& action) override { - TRACE_smpi_comm_in(my_proc_id, name, new simgrid::instr::CollTIData(name, (name == "gather") ? args.root : -1, -1.0, args.send_size, args.recv_size, + 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") + 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); @@ -501,291 +831,109 @@ public: TRACE_smpi_comm_out(my_proc_id); } }; -} // Replay Namespace -static void action_scatter(simgrid::xbt::ReplayAction& 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 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); - - TRACE_smpi_comm_out(Actor::self()->getPid()); - log_timed_action(action, clock); -} +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(); -static void action_gatherv(simgrid::xbt::ReplayAction& 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 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); - - int root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; - int rank = MPI_COMM_WORLD->rank(); - - if(rank==root) - recv = smpi_get_tmp_recvbuffer(recv_sum* MPI_CURRENT_TYPE2->size()); - - 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))); - - Colls::gatherv(send, send_size, MPI_CURRENT_TYPE, recv, recvcounts->data(), disps.data(), MPI_CURRENT_TYPE2, root, - MPI_COMM_WORLD); - - TRACE_smpi_comm_out(Actor::self()->getPid()); - log_timed_action (action, clock); -} + 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))); -static void action_scatterv(simgrid::xbt::ReplayAction& 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 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)); - - 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 = 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]); - } - 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()); - - 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))); - - Colls::scatterv(send, sendcounts->data(), disps.data(), 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); -} + 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); + } -static void action_reducescatter(simgrid::xbt::ReplayAction& 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 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; - - 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)}; - - 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))); - - 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->data(), MPI_CURRENT_TYPE, MPI_OP_NULL, MPI_COMM_WORLD); - smpi_execute_flops(comp_size); + TRACE_smpi_comm_out(my_proc_id); + } +}; - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); -} +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))); -static void action_allgatherv(simgrid::xbt::ReplayAction& 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 - 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]); - } - - 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}; - - void *sendbuf = smpi_get_tmp_sendbuffer(sendcount* MPI_CURRENT_TYPE->size()); - - 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()); - - int my_proc_id = Actor::self()->getPid(); - - 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))); - - Colls::allgatherv(sendbuf, sendcount, MPI_CURRENT_TYPE, recvbuf, recvcounts->data(), disps.data(), MPI_CURRENT_TYPE2, + 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); + } +}; + + +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))); + + 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); - log_timed_action (action, clock); -} + TRACE_smpi_comm_out(my_proc_id); + } +}; -static void action_allToAllv(simgrid::xbt::ReplayAction& 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(); - - 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))); - - Colls::alltoallv(sendbuf, sendcounts->data(), senddisps.data(), MPI_CURRENT_TYPE, recvbuf, recvcounts->data(), - recvdisps.data(), MPI_CURRENT_TYPE, MPI_COMM_WORLD); +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))); - TRACE_smpi_comm_out(my_proc_id); - log_timed_action (action, clock); -} + 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) { @@ -793,38 +941,38 @@ void smpi_replay_init(int* argc, char*** argv) smpi_process()->mark_as_initialized(); smpi_process()->set_replaying(true); - int my_proc_id = Actor::self()->getPid(); + int my_proc_id = simgrid::s4u::this_actor::get_pid(); 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("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").execute(action); }); - xbt_replay_action_register("Isend", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::SendAction("Isend").execute(action); }); - xbt_replay_action_register("recv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::RecvAction("recv").execute(action); }); - xbt_replay_action_register("Irecv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::RecvAction("Irecv").execute(action); }); - xbt_replay_action_register("test", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::TestAction().execute(action); }); - xbt_replay_action_register("wait", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::WaitAction().execute(action); }); - xbt_replay_action_register("waitAll", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::WaitAllAction().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::smpi::action_allToAllv); - xbt_replay_action_register("gather", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::GatherAction("gather").execute(action); }); - 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::xbt::ReplayAction& action) { simgrid::smpi::Replay::GatherAction("allGather").execute(action); }); - xbt_replay_action_register("allGatherV", simgrid::smpi::action_allgatherv); - xbt_replay_action_register("reduceScatter", simgrid::smpi::action_reducescatter); - xbt_replay_action_register("compute", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::Replay::ComputeAction().execute(action); }); + 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").execute(action); }); + xbt_replay_action_register("Isend", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SendAction("Isend").execute(action); }); + xbt_replay_action_register("recv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::RecvAction("recv").execute(action); }); + xbt_replay_action_register("Irecv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::RecvAction("Irecv").execute(action); }); + xbt_replay_action_register("test", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::TestAction().execute(action); }); + xbt_replay_action_register("wait", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::WaitAction().execute(action); }); + xbt_replay_action_register("waitAll", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::WaitAllAction().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){ @@ -841,6 +989,8 @@ 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 */ @@ -867,12 +1017,13 @@ 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()); + TRACE_smpi_finalize(simgrid::s4u::this_actor::get_pid()); } /** @brief chain a replay initialization and a replay start */