X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/970f6b86abcc928064341baf688313fc0a051660..de190af4974a9ed57a52e649e5cbe0ac83512acb:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index de586295a7..8459a7f268 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -19,11 +19,51 @@ #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 class hash { +public: + 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> class HashValueImpl { +public: + static void apply(size_t& seed, Tuple const& tuple) + { + HashValueImpl::apply(seed, tuple); + hash_combine(seed, std::get(tuple)); + } +}; + +template class HashValueImpl { +public: + static void apply(size_t& seed, Tuple const& tuple) { hash_combine(seed, std::get<0>(tuple)); } +}; + +template class hash> { +public: + 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 std::unordered_map*> reqq; +typedef std::tuple req_key_t; +typedef std::unordered_map>> req_storage_t; static MPI_Datatype MPI_DEFAULT_TYPE; @@ -52,16 +92,6 @@ static void log_timed_action(simgrid::xbt::ReplayAction& action, double clock) } } -static std::vector* get_reqq_self() -{ - return reqq.at(simgrid::s4u::this_actor::get_pid()); -} - -static void set_reqq_self(std::vector *mpi_request) -{ - reqq.insert({simgrid::s4u::this_actor::get_pid(), mpi_request}); -} - /* Helper function */ static double parse_double(std::string string) { @@ -72,12 +102,85 @@ namespace simgrid { namespace smpi { 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::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"); + } + } + } + + 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}); + } +}; + +/** + * Base class for all parsers. + */ 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]); + } +}; + class SendRecvParser : public ActionArgParser { public: /* communication partner; if we send, this is the receiver and vice versa */ @@ -400,6 +503,12 @@ public: } }; +/** + * Base class for all ReplayActions. + * Note that this class actually implements the behavior of each action + * while the parsing of the replay arguments is done in the @ActionArgParser class. + * In other words: The logic goes here, the setup is done by the ActionArgParser. + */ template class ReplayAction { protected: const std::string name; @@ -433,17 +542,20 @@ public: } }; -class WaitAction : public ReplayAction { +class WaitAction : public ReplayAction { +private: + RequestStorage& req_storage; + public: - WaitAction() : ReplayAction("Wait") {} + 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(get_reqq_self()->size(), "action wait not preceded by any irecv or isend: %s", s.c_str()); - MPI_Request request = get_reqq_self()->back(); - get_reqq_self()->pop_back(); + 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 == 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; @@ -453,9 +565,6 @@ 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()); - int tag = request->tag(); 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")); @@ -465,14 +574,16 @@ public: TRACE_smpi_comm_out(rank); if (is_wait_for_receive) - TRACE_smpi_recv(src, dst, tag); + TRACE_smpi_recv(args.src, args.dst, args.tag); } }; class SendAction : public ReplayAction { +private: + RequestStorage& req_storage; + public: - SendAction() = delete; - explicit SendAction(std::string name) : ReplayAction(name) {} + explicit SendAction(std::string name, RequestStorage& storage) : ReplayAction(name), req_storage(storage) {} void kernel(simgrid::xbt::ReplayAction& action) override { int dst_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); @@ -486,7 +597,7 @@ public: 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); - get_reqq_self()->push_back(request); + req_storage.add(request); } else { xbt_die("Don't know this action, %s", name.c_str()); } @@ -496,9 +607,11 @@ public: }; class RecvAction : public ReplayAction { +private: + RequestStorage& req_storage; + public: - RecvAction() = delete; - explicit RecvAction(std::string name) : ReplayAction(name) {} + explicit RecvAction(std::string name, RequestStorage& storage) : ReplayAction(name), req_storage(storage) {} void kernel(simgrid::xbt::ReplayAction& action) override { int src_traced = MPI_COMM_WORLD->group()->actor(args.partner)->get_pid(); @@ -517,7 +630,7 @@ public: 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); - get_reqq_self()->push_back(request); + req_storage.add(request); } TRACE_smpi_comm_out(my_proc_id); @@ -539,17 +652,20 @@ public: } }; -class TestAction : public ReplayAction { +class TestAction : public ReplayAction { +private: + RequestStorage& req_storage; + public: - TestAction() : ReplayAction("Test") {} + explicit TestAction(RequestStorage& storage) : ReplayAction("Test"), req_storage(storage) {} void kernel(simgrid::xbt::ReplayAction& action) override { - MPI_Request request = get_reqq_self()->back(); - get_reqq_self()->pop_back(); + 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 != nullptr) { + if (request != MPI_REQUEST_NULL) { TRACE_smpi_testing_in(my_proc_id); MPI_Status status; @@ -558,7 +674,10 @@ public: 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); + 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); } @@ -576,7 +695,6 @@ public: /* start a simulated timer */ smpi_process()->simulated_start(); - set_reqq_self(new std::vector); } }; @@ -587,22 +705,28 @@ public: }; class WaitAllAction : public ReplayAction { +private: + RequestStorage& req_storage; + public: - WaitAllAction() : ReplayAction("waitAll") {} + explicit WaitAllAction(RequestStorage& storage) : ReplayAction("waitAll"), req_storage(storage) {} void kernel(simgrid::xbt::ReplayAction& action) override { - const unsigned int count_requests = get_reqq_self()->size(); + 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; - for (const auto& req : (*get_reqq_self())) { + 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, &(*get_reqq_self())[0], status); + Request::waitall(count_requests, &(reqs.data())[0], status); + req_storage.get_store().clear(); for (auto& pair : sender_receiver) { TRACE_smpi_recv(pair.first, pair.second, 0); @@ -812,6 +936,7 @@ public: } // 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) { @@ -820,6 +945,8 @@ void smpi_replay_init(int* argc, char*** argv) smpi_process()->set_replaying(true); 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")); @@ -829,14 +956,13 @@ void smpi_replay_init(int* argc, char*** argv) 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("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); }); @@ -873,20 +999,19 @@ void smpi_replay_main(int* argc, char*** 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){