#include "smpi_process.hpp"
#include "smpi_request.hpp"
#include "xbt/replay.hpp"
+#include <simgrid/smpi/replay.hpp>
#include <boost/algorithm/string/join.hpp>
#include <memory>
#include <numeric>
#include <unordered_map>
-#include <sstream>
#include <vector>
#include <tuple>
// 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 <typename TT>
- class hash
- {
- public:
- size_t
- operator()(TT const& tt) const
- {
- return std::hash<TT>()(tt);
- }
- };
+namespace hash_tuple {
+template <typename TT> class hash {
+public:
+ size_t operator()(TT const& tt) const { return std::hash<TT>()(tt); }
+};
- template <class T>
- inline void hash_combine(std::size_t& seed, T const& v)
- {
- seed ^= hash_tuple::hash<T>()(v) + 0x9e3779b9 + (seed<<6) + (seed>>2);
- }
+template <class T> inline void hash_combine(std::size_t& seed, T const& v)
+{
+ seed ^= hash_tuple::hash<T>()(v) + 0x9e3779b9 + (seed << 6) + (seed >> 2);
+}
- // Recursive template code derived from Matthieu M.
- template <class Tuple, size_t Index = std::tuple_size<Tuple>::value - 1>
- class HashValueImpl
- {
- public:
- static void apply(size_t& seed, Tuple const& tuple)
- {
- HashValueImpl<Tuple, Index-1>::apply(seed, tuple);
- hash_combine(seed, std::get<Index>(tuple));
- }
- };
+// Recursive template code derived from Matthieu M.
+template <class Tuple, size_t Index = std::tuple_size<Tuple>::value - 1> class HashValueImpl {
+public:
+ static void apply(size_t& seed, Tuple const& tuple)
+ {
+ HashValueImpl<Tuple, Index - 1>::apply(seed, tuple);
+ hash_combine(seed, std::get<Index>(tuple));
+ }
+};
- template <class Tuple>
- class HashValueImpl<Tuple,0>
- {
- public:
- static void apply(size_t& seed, Tuple const& tuple)
- {
- hash_combine(seed, std::get<0>(tuple));
- }
- };
+template <class Tuple> class HashValueImpl<Tuple, 0> {
+public:
+ static void apply(size_t& seed, Tuple const& tuple) { hash_combine(seed, std::get<0>(tuple)); }
+};
- template <typename ... TT>
- class hash<std::tuple<TT...>>
- {
- public:
- size_t
- operator()(std::tuple<TT...> const& tt) const
- {
- size_t seed = 0;
- HashValueImpl<std::tuple<TT...> >::apply(seed, tt);
- return seed;
- }
- };
+template <typename... TT> class hash<std::tuple<TT...>> {
+public:
+ size_t operator()(std::tuple<TT...> const& tt) const
+ {
+ size_t seed = 0;
+ HashValueImpl<std::tuple<TT...>>::apply(seed, tt);
+ return seed;
+ }
+};
}
XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_replay,smpi,"Trace Replay with SMPI");
typedef std::tuple</*sender*/ int, /* reciever */ int, /* tag */int> req_key_t;
typedef std::unordered_map<req_key_t, MPI_Request, hash_tuple::hash<std::tuple<int,int,int>>> req_storage_t;
-static MPI_Datatype MPI_DEFAULT_TYPE;
-
-#define CHECK_ACTION_PARAMS(action, mandatory, optional) \
- { \
- if (action.size() < static_cast<unsigned long>(mandatory + 2)) { \
- std::stringstream ss; \
- for (const auto& elem : action) { \
- ss << elem << " "; \
- } \
- THROWF(arg_error, 0, "%s replay failed.\n" \
- "%zu items were given on the line. First two should be process_id and action. " \
- "This action needs after them %lu mandatory arguments, and accepts %lu optional ones. \n" \
- "The full line that was given is:\n %s\n" \
- "Please contact the Simgrid team if support is needed", \
- __func__, action.size(), static_cast<unsigned long>(mandatory), static_cast<unsigned long>(optional), \
- ss.str().c_str()); \
- } \
- }
static void log_timed_action(simgrid::xbt::ReplayAction& action, double clock)
{
namespace smpi {
namespace replay {
+MPI_Datatype MPI_DEFAULT_TYPE;
class RequestStorage {
private:
}
};
-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
+ void WaitTestParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
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 */
- int partner;
- double size;
- int tag;
- MPI_Datatype datatype1 = MPI_DEFAULT_TYPE;
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void SendRecvParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
CHECK_ACTION_PARAMS(action, 3, 1)
partner = std::stoi(action[2]);
if (action.size() > 5)
datatype1 = simgrid::smpi::Datatype::decode(action[5]);
}
-};
-class ComputeParser : public ActionArgParser {
-public:
- /* communication partner; if we send, this is the receiver and vice versa */
- double flops;
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void ComputeParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
CHECK_ACTION_PARAMS(action, 1, 0)
flops = parse_double(action[2]);
}
-};
-
-class CollCommParser : public ActionArgParser {
-public:
- double size;
- double comm_size;
- double comp_size;
- int send_size;
- int recv_size;
- int root = 0;
- MPI_Datatype datatype1 = MPI_DEFAULT_TYPE;
- MPI_Datatype datatype2 = MPI_DEFAULT_TYPE;
-};
-class BcastArgParser : public CollCommParser {
-public:
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void BcastArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
CHECK_ACTION_PARAMS(action, 1, 2)
size = parse_double(action[2]);
if (action.size() > 4)
datatype1 = simgrid::smpi::Datatype::decode(action[4]);
}
-};
-class ReduceArgParser : public CollCommParser {
-public:
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void ReduceArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
CHECK_ACTION_PARAMS(action, 2, 2)
comm_size = parse_double(action[2]);
if (action.size() > 5)
datatype1 = simgrid::smpi::Datatype::decode(action[5]);
}
-};
-class AllReduceArgParser : public CollCommParser {
-public:
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void AllReduceArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
CHECK_ACTION_PARAMS(action, 2, 1)
comm_size = parse_double(action[2]);
if (action.size() > 4)
datatype1 = simgrid::smpi::Datatype::decode(action[4]);
}
-};
-class AllToAllArgParser : public CollCommParser {
-public:
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void AllToAllArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
CHECK_ACTION_PARAMS(action, 2, 1)
comm_size = MPI_COMM_WORLD->size();
if (action.size() > 5)
datatype2 = simgrid::smpi::Datatype::decode(action[5]);
}
-};
-class GatherArgParser : public CollCommParser {
-public:
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void GatherArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
/* The structure of the gather action for the rank 0 (total 4 processes) is the following:
0 gather 68 68 0 0 0
datatype2 = simgrid::smpi::Datatype::decode(action[5]);
}
}
-};
-class GatherVArgParser : public CollCommParser {
-public:
- int recv_size_sum;
- std::shared_ptr<std::vector<int>> recvcounts;
- std::vector<int> disps;
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void GatherVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
/* 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
}
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
+ void ScatterArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
/* The structure of the scatter action for the rank 0 (total 4 processes) is the following:
0 gather 68 68 0 0 0
if (action.size() > 6)
datatype2 = simgrid::smpi::Datatype::decode(action[6]);
}
-};
-class ScatterVArgParser : public CollCommParser {
-public:
- int recv_size_sum;
- int send_size_sum;
- std::shared_ptr<std::vector<int>> sendcounts;
- std::vector<int> disps;
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void ScatterVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
/* 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
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<std::vector<int>> recvcounts;
- std::vector<int> disps;
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void ReduceScatterArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
/* The structure of the reducescatter action for the rank 0 (total 4 processes) is the following:
0 reduceScatter 275427 275427 275427 204020 11346849 0
}
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<std::vector<int>> recvcounts;
- std::shared_ptr<std::vector<int>> sendcounts;
- std::vector<int> senddisps;
- std::vector<int> recvdisps;
- int send_buf_size;
- int recv_buf_size;
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
+ void AllToAllVArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
{
/* 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
send_size_sum = std::accumulate(sendcounts->begin(), sendcounts->end(), 0);
recv_size_sum = std::accumulate(recvcounts->begin(), recvcounts->end(), 0);
}
-};
-
-template <class T> class ReplayAction {
-protected:
- const std::string name;
- RequestStorage* req_storage; // Points to the right storage for this process, nullptr except for Send/Recv/Wait/Test actions.
- const int my_proc_id;
- T args;
-
-public:
- explicit ReplayAction(std::string name, RequestStorage& storage) : name(name), req_storage(&storage), my_proc_id(simgrid::s4u::this_actor::get_pid()) {}
- explicit ReplayAction(std::string name) : name(name), req_storage(nullptr), my_proc_id(simgrid::s4u::this_actor::get_pid()) {}
- virtual ~ReplayAction() = default;
-
- virtual void execute(simgrid::xbt::ReplayAction& action)
- {
- // Needs to be re-initialized for every action, hence here
- double start_time = smpi_process()->simulated_elapsed();
- args.parse(action, name);
- kernel(action);
- if (name != "Init")
- log_timed_action(action, start_time);
- }
-
- virtual void kernel(simgrid::xbt::ReplayAction& action) = 0;
-
- void* send_buffer(int size)
- {
- return smpi_get_tmp_sendbuffer(size);
- }
- void* recv_buffer(int size)
- {
- return smpi_get_tmp_recvbuffer(size);
- }
-};
+template<class T>
+void ReplayAction<T>::execute(simgrid::xbt::ReplayAction& action)
+{
+ // Needs to be re-initialized for every action, hence here
+ double start_time = smpi_process()->simulated_elapsed();
+ args.parse(action, name);
+ kernel(action);
+ if (name != "Init")
+ log_timed_action(action, start_time);
+}
class WaitAction : public ReplayAction<WaitTestParser> {
+private:
+ RequestStorage& req_storage;
+
public:
- explicit WaitAction(RequestStorage& storage) : ReplayAction("Wait", storage) {}
+ 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);
+ 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
};
class SendAction : public ReplayAction<SendRecvParser> {
+private:
+ RequestStorage& req_storage;
+
public:
- SendAction() = delete;
- explicit SendAction(std::string name, RequestStorage& storage) : ReplayAction(name, storage) {}
+ 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();
Request::send(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD);
} else if (name == "Isend") {
MPI_Request request = Request::isend(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD);
- req_storage->add(request);
+ req_storage.add(request);
} else {
xbt_die("Don't know this action, %s", name.c_str());
}
};
class RecvAction : public ReplayAction<SendRecvParser> {
+private:
+ RequestStorage& req_storage;
+
public:
- RecvAction() = delete;
- explicit RecvAction(std::string name, RequestStorage& storage) : ReplayAction(name, storage) {}
+ 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();
Request::recv(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD, &status);
} else if (name == "Irecv") {
MPI_Request request = Request::irecv(nullptr, args.size, args.datatype1, args.partner, args.tag, MPI_COMM_WORLD);
- req_storage->add(request);
+ req_storage.add(request);
}
TRACE_smpi_comm_out(my_proc_id);
};
class TestAction : public ReplayAction<WaitTestParser> {
+private:
+ RequestStorage& req_storage;
+
public:
- explicit TestAction(RequestStorage& storage) : ReplayAction("Test", storage) {}
+ explicit TestAction(RequestStorage& storage) : ReplayAction("Test"), req_storage(storage) {}
void kernel(simgrid::xbt::ReplayAction& action) override
{
- MPI_Request request = req_storage->find(args.src, args.dst, args.tag);
- req_storage->remove(request);
+ 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.
/* push back request in vector to be caught by a subsequent wait. if the test did succeed, the request is now
* nullptr.*/
if (request == MPI_REQUEST_NULL)
- req_storage->addNullRequest(args.src, args.dst, args.tag);
+ req_storage.addNullRequest(args.src, args.dst, args.tag);
else
- req_storage->add(request);
+ req_storage.add(request);
TRACE_smpi_testing_out(my_proc_id);
}
};
class WaitAllAction : public ReplayAction<ActionArgParser> {
+private:
+ RequestStorage& req_storage;
+
public:
- explicit WaitAllAction(RequestStorage& storage) : ReplayAction("waitAll", storage) {}
+ explicit WaitAllAction(RequestStorage& storage) : ReplayAction("waitAll"), req_storage(storage) {}
void kernel(simgrid::xbt::ReplayAction& action) override
{
- const unsigned int count_requests = req_storage->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<std::pair</*sender*/int,/*recv*/int>> sender_receiver;
std::vector<MPI_Request> reqs;
- req_storage->get_requests(reqs);
+ req_storage.get_requests(reqs);
for (const auto& req : reqs) {
if (req && (req->flags() & RECV)) {
sender_receiver.push_back({req->src(), req->dst()});
}
MPI_Status status[count_requests];
Request::waitall(count_requests, &(reqs.data())[0], status);
- req_storage->get_store().clear();
+ req_storage.get_store().clear();
for (auto& pair : sender_receiver) {
TRACE_smpi_recv(pair.first, pair.second, 0);
xbt_replay_action_register("comm_size", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::CommunicatorAction().execute(action); });
xbt_replay_action_register("comm_split",[](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::CommunicatorAction().execute(action); });
xbt_replay_action_register("comm_dup", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::CommunicatorAction().execute(action); });
-
xbt_replay_action_register("send", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SendAction("send", storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); });
xbt_replay_action_register("Isend", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::SendAction("Isend", storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); });
xbt_replay_action_register("recv", [](simgrid::xbt::ReplayAction& action) { simgrid::smpi::replay::RecvAction("recv", storage[simgrid::s4u::this_actor::get_pid()-1]).execute(action); });
