X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/4d4fe07327faa75ebe7dd12ff343b961fe04cb51..4efcdc1e4b35fbb46a705f163bc5c1745ec342c9:/src/smpi/internals/smpi_replay.cpp?ds=sidebyside diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index 7fc301d672..d2727d87be 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -3,23 +3,24 @@ /* This program is free software; you can redistribute it and/or modify it * under the terms of the license (GNU LGPL) which comes with this package. */ -#include "private.hpp" #include "smpi_coll.hpp" #include "smpi_comm.hpp" #include "smpi_datatype.hpp" #include "smpi_group.hpp" -#include "smpi_process.hpp" #include "smpi_request.hpp" #include "xbt/replay.hpp" #include +#include -#include #include #include #include #include #include + +XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay, smpi, "Trace Replay with SMPI"); + // From https://stackoverflow.com/questions/7110301/generic-hash-for-tuples-in-unordered-map-unordered-set // This is all just to make std::unordered_map work with std::tuple. If we need this in other places, // this could go into a header file. @@ -60,8 +61,6 @@ public: }; } -XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI"); - typedef std::tuple req_key_t; typedef std::unordered_map>> req_storage_t; @@ -158,7 +157,6 @@ void SendRecvParser::parse(simgrid::xbt::ReplayAction& action, std::string name) datatype1 = simgrid::smpi::Datatype::decode(action[5]); } - void ComputeParser::parse(simgrid::xbt::ReplayAction& action, std::string name) { CHECK_ACTION_PARAMS(action, 1, 0) @@ -228,8 +226,7 @@ void GatherArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name datatype1 = simgrid::smpi::Datatype::decode(action[5]); if (action.size() > 6) datatype2 = simgrid::smpi::Datatype::decode(action[6]); - } - else { + } else { if (action.size() > 4) datatype1 = simgrid::smpi::Datatype::decode(action[4]); if (action.size() > 5) @@ -249,9 +246,9 @@ void GatherVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string nam 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); + 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") { @@ -260,8 +257,7 @@ void GatherVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string nam 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 { + } else { int datatype_index = 0; int disp_index = 0; /* The 3 comes from "0 gather ", which must always be present. @@ -272,9 +268,10 @@ void GatherVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string nam 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 */ + } 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]); @@ -304,10 +301,10 @@ void ScatterArgParser::parse(simgrid::xbt::ReplayAction& action, std::string nam 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; + 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) @@ -339,7 +336,7 @@ void ScatterVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string na (*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; + root = (action.size() > 3 + comm_size) ? std::stoi(action[3 + comm_size]) : 0; } void ReduceScatterArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name) @@ -352,8 +349,8 @@ void ReduceScatterArgParser::parse(simgrid::xbt::ReplayAction& action, std::stri 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]); + 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]); @@ -375,7 +372,7 @@ void AllToAllVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string n 4) 1 70 10 5 is the recvcounts array */ comm_size = MPI_COMM_WORLD->size(); - CHECK_ACTION_PARAMS(action, 2*comm_size+2, 2) + 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); @@ -386,8 +383,8 @@ void AllToAllVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string n 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]); + 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]); @@ -407,49 +404,36 @@ void ReplayAction::execute(simgrid::xbt::ReplayAction& action) log_timed_action(action, start_time); } -class WaitAction : public ReplayAction { -private: - RequestStorage& req_storage; - -public: - explicit WaitAction(RequestStorage& storage) : ReplayAction("Wait"), req_storage(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); +void WaitAction::kernel(simgrid::xbt::ReplayAction& action) +{ + std::string s = boost::algorithm::join(action, " "); + xbt_assert(req_storage.size(), "action wait not preceded by any irecv or isend: %s", s.c_str()); + 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 (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; + } - int rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1; + int rank = request->comm() != MPI_COMM_NULL ? request->comm()->rank() : -1; - // Must be taken before Request::wait() since the request may be set to - // MPI_REQUEST_NULL by Request::wait! - bool is_wait_for_receive = (request->flags() & 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")); + // Must be taken before Request::wait() since the request may be set to + // MPI_REQUEST_NULL by Request::wait! + bool is_wait_for_receive = (request->flags() & MPI_REQ_RECV); + // TODO: Here we take the rank while we normally take the process id (look for my_proc_id) + TRACE_smpi_comm_in(rank, __func__, new simgrid::instr::WaitTIData(args.src, args.dst, args.tag)); - MPI_Status status; - Request::wait(&request, &status); + MPI_Status status; + Request::wait(&request, &status); - TRACE_smpi_comm_out(rank); - if (is_wait_for_receive) - TRACE_smpi_recv(args.src, args.dst, args.tag); + TRACE_smpi_comm_out(rank); + if (is_wait_for_receive) + TRACE_smpi_recv(args.src, args.dst, args.tag); } -}; - -class SendAction : public ReplayAction { -private: - RequestStorage& req_storage; -public: - explicit SendAction(std::string name, RequestStorage& storage) : ReplayAction(name), req_storage(storage) {} - void kernel(simgrid::xbt::ReplayAction& action) override + void SendAction::kernel(simgrid::xbt::ReplayAction& action) { int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); @@ -469,15 +453,8 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; -class RecvAction : public ReplayAction { -private: - RequestStorage& req_storage; - -public: - explicit RecvAction(std::string name, RequestStorage& storage) : ReplayAction(name), req_storage(storage) {} - void kernel(simgrid::xbt::ReplayAction& action) override + void RecvAction::kernel(simgrid::xbt::ReplayAction& action) { int src_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); @@ -504,26 +481,15 @@ public: 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 + void ComputeAction::kernel(simgrid::xbt::ReplayAction& action) { 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 { -private: - RequestStorage& req_storage; - -public: - explicit TestAction(RequestStorage& storage) : ReplayAction("Test"), req_storage(storage) {} - void kernel(simgrid::xbt::ReplayAction& action) override + void TestAction::kernel(simgrid::xbt::ReplayAction& action) { MPI_Request request = req_storage.find(args.src, args.dst, args.tag); req_storage.remove(request); @@ -531,7 +497,7 @@ public: // 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); + TRACE_smpi_comm_in(my_proc_id, __func__, new simgrid::instr::NoOpTIData("test")); MPI_Status status; int flag = Request::test(&request, &status); @@ -544,15 +510,11 @@ public: else req_storage.add(request); - TRACE_smpi_testing_out(my_proc_id); + TRACE_smpi_comm_out(my_proc_id); } } -}; -class InitAction : public ReplayAction { -public: - InitAction() : ReplayAction("Init") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void InitAction::kernel(simgrid::xbt::ReplayAction& action) { CHECK_ACTION_PARAMS(action, 0, 1) MPI_DEFAULT_TYPE = (action.size() > 2) ? MPI_DOUBLE // default MPE datatype @@ -561,21 +523,13 @@ public: /* 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 { -private: - RequestStorage& req_storage; + void CommunicatorAction::kernel(simgrid::xbt::ReplayAction& action) + { + /* nothing to do */ + } -public: - explicit WaitAllAction(RequestStorage& storage) : ReplayAction("waitAll"), req_storage(storage) {} - void kernel(simgrid::xbt::ReplayAction& action) override + void WaitAllAction::kernel(simgrid::xbt::ReplayAction& action) { const unsigned int count_requests = req_storage.size(); @@ -585,7 +539,7 @@ public: std::vector reqs; req_storage.get_requests(reqs); for (const auto& req : reqs) { - if (req && (req->flags() & RECV)) { + if (req && (req->flags() & MPI_REQ_RECV)) { sender_receiver.push_back({req->src(), req->dst()}); } } @@ -599,23 +553,15 @@ public: TRACE_smpi_comm_out(my_proc_id); } } -}; -class BarrierAction : public ReplayAction { -public: - BarrierAction() : ReplayAction("barrier") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void BarrierAction::kernel(simgrid::xbt::ReplayAction& action) { 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); } -}; -class BcastAction : public ReplayAction { -public: - BcastAction() : ReplayAction("bcast") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void BcastAction::kernel(simgrid::xbt::ReplayAction& action) { TRACE_smpi_comm_in(my_proc_id, "action_bcast", new simgrid::instr::CollTIData("bcast", MPI_COMM_WORLD->group()->actor(args.root)->get_pid(), @@ -625,12 +571,8 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; -class ReduceAction : public ReplayAction { -public: - ReduceAction() : ReplayAction("reduce") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void ReduceAction::kernel(simgrid::xbt::ReplayAction& action) { TRACE_smpi_comm_in(my_proc_id, "action_reduce", new simgrid::instr::CollTIData("reduce", MPI_COMM_WORLD->group()->actor(args.root)->get_pid(), @@ -643,12 +585,8 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; -class AllReduceAction : public ReplayAction { -public: - AllReduceAction() : ReplayAction("allReduce") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void AllReduceAction::kernel(simgrid::xbt::ReplayAction& action) { 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), "")); @@ -659,12 +597,8 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; -class AllToAllAction : public ReplayAction { -public: - AllToAllAction() : ReplayAction("allToAll") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void AllToAllAction::kernel(simgrid::xbt::ReplayAction& action) { TRACE_smpi_comm_in(my_proc_id, "action_allToAll", new simgrid::instr::CollTIData("allToAll", -1, -1.0, args.send_size, args.recv_size, @@ -677,12 +611,8 @@ public: 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 + void GatherAction::kernel(simgrid::xbt::ReplayAction& action) { 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))); @@ -698,12 +628,8 @@ public: 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 + void GatherVAction::kernel(simgrid::xbt::ReplayAction& action) { int rank = MPI_COMM_WORLD->rank(); @@ -724,12 +650,8 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; -class ScatterAction : public ReplayAction { -public: - ScatterAction() : ReplayAction("scatter") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void ScatterAction::kernel(simgrid::xbt::ReplayAction& action) { 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, @@ -741,13 +663,8 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; - -class ScatterVAction : public ReplayAction { -public: - ScatterVAction() : ReplayAction("scatterV") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void ScatterVAction::kernel(simgrid::xbt::ReplayAction& action) { 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, @@ -761,12 +678,8 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; -class ReduceScatterAction : public ReplayAction { -public: - ReduceScatterAction() : ReplayAction("reduceScatter") {} - void kernel(simgrid::xbt::ReplayAction& action) override + void ReduceScatterAction::kernel(simgrid::xbt::ReplayAction& action) { TRACE_smpi_comm_in(my_proc_id, "action_reducescatter", new simgrid::instr::VarCollTIData("reduceScatter", -1, 0, nullptr, -1, args.recvcounts, @@ -780,12 +693,8 @@ public: 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 + void AllToAllVAction::kernel(simgrid::xbt::ReplayAction& action) { TRACE_smpi_comm_in(my_proc_id, __func__, new simgrid::instr::VarCollTIData( @@ -797,11 +706,10 @@ public: TRACE_smpi_comm_out(my_proc_id); } -}; } // Replay Namespace }} // namespace simgrid::smpi -std::vector storage; +static std::vector storage; /** @brief Only initialize the replay, don't do it for real */ void smpi_replay_init(int* argc, char*** argv) {