X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/970f6b86abcc928064341baf688313fc0a051660..8fc62f8b40bf977a42f16281950ce7564cd6f91a:/src/smpi/internals/smpi_replay.cpp diff --git a/src/smpi/internals/smpi_replay.cpp b/src/smpi/internals/smpi_replay.cpp index de586295a7..e7c7a07a15 100644 --- a/src/smpi/internals/smpi_replay.cpp +++ b/src/smpi/internals/smpi_replay.cpp @@ -21,9 +21,65 @@ 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 std::unordered_map*> reqq; +typedef std::tuple req_key_t; +typedef std::unordered_map>> req_storage_t; static MPI_Datatype MPI_DEFAULT_TYPE; @@ -72,12 +128,82 @@ 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}); + } +}; + 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 */ @@ -403,11 +529,13 @@ public: 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) : name(name), my_proc_id(simgrid::s4u::this_actor::get_pid()) {} + 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) @@ -433,17 +561,17 @@ public: } }; -class WaitAction : public ReplayAction { +class WaitAction : public ReplayAction { public: - WaitAction() : ReplayAction("Wait") {} + WaitAction(RequestStorage& storage) : ReplayAction("Wait", 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 +581,6 @@ 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 +590,14 @@ 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 { public: SendAction() = delete; - explicit SendAction(std::string name) : ReplayAction(name) {} + 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(); @@ -498,7 +623,7 @@ public: class RecvAction : public ReplayAction { public: RecvAction() = delete; - explicit RecvAction(std::string name) : ReplayAction(name) {} + 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(); @@ -539,9 +664,9 @@ public: } }; -class TestAction : public ReplayAction { +class TestAction : public ReplayAction { public: - TestAction() : ReplayAction("Test") {} + TestAction(RequestStorage& storage) : ReplayAction("Test", storage) {} void kernel(simgrid::xbt::ReplayAction& action) override { MPI_Request request = get_reqq_self()->back(); @@ -549,7 +674,7 @@ public: // 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; @@ -588,7 +713,7 @@ public: class WaitAllAction : public ReplayAction { public: - WaitAllAction() : ReplayAction("waitAll") {} + WaitAllAction(RequestStorage& storage) : ReplayAction("waitAll", storage) {} void kernel(simgrid::xbt::ReplayAction& action) override { const unsigned int count_requests = get_reqq_self()->size(); @@ -812,6 +937,7 @@ 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 +946,8 @@ 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")); @@ -830,13 +958,13 @@ void smpi_replay_init(int* argc, char*** argv) 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); });