#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>
-using simgrid::s4u::Actor;
-
#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>
- struct hash
- {
- 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>
- struct HashValueImpl
- {
- 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>
- struct HashValueImpl<Tuple,0>
- {
- 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>
- struct hash<std::tuple<TT...>>
- {
- 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");
-static std::unordered_map<int, std::vector<MPI_Request>*> reqq;
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)
{
}
}
-static std::vector<MPI_Request>* get_reqq_self()
-{
- return reqq.at(simgrid::s4u::this_actor::get_pid());
-}
-
-static void set_reqq_self(std::vector<MPI_Request> *mpi_request)
-{
- reqq.insert({simgrid::s4u::this_actor::get_pid(), mpi_request});
-}
-
/* Helper function */
static double parse_double(std::string string)
{
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 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]);
+}
- 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]);
- }
-};
+void SendRecvParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
+{
+ CHECK_ACTION_PARAMS(action, 3, 1)
+ partner = std::stoi(action[2]);
+ tag = std::stoi(action[3]);
+ size = parse_double(action[4]);
+ if (action.size() > 5)
+ datatype1 = simgrid::smpi::Datatype::decode(action[5]);
+}
-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 ComputeParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
+{
+ CHECK_ACTION_PARAMS(action, 1, 0)
+ flops = parse_double(action[2]);
+}
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
- {
- CHECK_ACTION_PARAMS(action, 3, 1)
- partner = std::stoi(action[2]);
- tag = std::stoi(action[3]);
- size = parse_double(action[4]);
- if (action.size() > 5)
- datatype1 = simgrid::smpi::Datatype::decode(action[5]);
- }
-};
+void BcastArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
+{
+ CHECK_ACTION_PARAMS(action, 1, 2)
+ size = parse_double(action[2]);
+ root = (action.size() > 3) ? std::stoi(action[3]) : 0;
+ if (action.size() > 4)
+ datatype1 = simgrid::smpi::Datatype::decode(action[4]);
+}
-class ComputeParser : public ActionArgParser {
-public:
- /* communication partner; if we send, this is the receiver and vice versa */
- double flops;
+void ReduceArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
+{
+ CHECK_ACTION_PARAMS(action, 2, 2)
+ comm_size = parse_double(action[2]);
+ comp_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]);
+}
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
- {
- CHECK_ACTION_PARAMS(action, 1, 0)
- flops = parse_double(action[2]);
- }
-};
+void AllReduceArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
+{
+ CHECK_ACTION_PARAMS(action, 2, 1)
+ comm_size = parse_double(action[2]);
+ comp_size = parse_double(action[3]);
+ if (action.size() > 4)
+ datatype1 = simgrid::smpi::Datatype::decode(action[4]);
+}
-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;
-};
+void AllToAllArgParser::parse(simgrid::xbt::ReplayAction& action, std::string name)
+{
+ CHECK_ACTION_PARAMS(action, 2, 1)
+ comm_size = MPI_COMM_WORLD->size();
+ send_size = parse_double(action[2]);
+ recv_size = parse_double(action[3]);
+
+ if (action.size() > 4)
+ datatype1 = simgrid::smpi::Datatype::decode(action[4]);
+ if (action.size() > 5)
+ datatype2 = simgrid::smpi::Datatype::decode(action[5]);
+}
-class BcastArgParser : public CollCommParser {
-public:
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
- {
- CHECK_ACTION_PARAMS(action, 1, 2)
- size = parse_double(action[2]);
- root = (action.size() > 3) ? std::stoi(action[3]) : 0;
- if (action.size() > 4)
- datatype1 = simgrid::smpi::Datatype::decode(action[4]);
- }
-};
+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
+ 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]);
-class ReduceArgParser : public CollCommParser {
-public:
- void parse(simgrid::xbt::ReplayAction& action, std::string name) override
- {
- CHECK_ACTION_PARAMS(action, 2, 2)
- comm_size = parse_double(action[2]);
- comp_size = parse_double(action[3]);
+ if (name == "gather") {
root = (action.size() > 4) ? std::stoi(action[4]) : 0;
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
- {
- CHECK_ACTION_PARAMS(action, 2, 1)
- comm_size = parse_double(action[2]);
- comp_size = parse_double(action[3]);
- 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
- {
- CHECK_ACTION_PARAMS(action, 2, 1)
- comm_size = MPI_COMM_WORLD->size();
- send_size = parse_double(action[2]);
- recv_size = parse_double(action[3]);
-
+ 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 GatherArgParser : public CollCommParser {
-public:
- 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
- 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]);
-
- 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<std::vector<int>> recvcounts;
- std::vector<int> 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<int>(comm_size, 0);
- recvcounts = std::shared_ptr<std::vector<int>>(new std::vector<int>(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 <sendcount>", 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]);
- }
+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
+ 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<int>(comm_size, 0);
+ recvcounts = std::shared_ptr<std::vector<int>>(new std::vector<int>(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 <sendcount>", 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]);
}
- for (unsigned int i = 0; i < comm_size; i++) {
- (*recvcounts)[i] = std::stoi(action[i + 3]);
+ if (disp_index != 0) {
+ for (unsigned int i = 0; i < comm_size; i++)
+ disps[i] = std::stoi(action[disp_index + i]);
}
- 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)
- datatype2 = simgrid::smpi::Datatype::decode(action[6]);
+ 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 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
- {
- /* 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
+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
+ 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, comm_size + 1, 2)
- recv_size = parse_double(action[2 + comm_size]);
- disps = std::vector<int>(comm_size, 0);
- sendcounts = std::shared_ptr<std::vector<int>>(new std::vector<int>(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<std::vector<int>> recvcounts;
- std::vector<int> 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<std::vector<int>>(new std::vector<int>(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<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
- {
- /* 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<std::vector<int>>(new std::vector<int>(comm_size));
- recvcounts = std::shared_ptr<std::vector<int>>(new std::vector<int>(comm_size));
- senddisps = std::vector<int>(comm_size, 0);
- recvdisps = std::vector<int>(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 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;
+ */
+ 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)
+ datatype2 = simgrid::smpi::Datatype::decode(action[6]);
+}
- 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);
- }
+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
+ 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<int>(comm_size, 0);
+ sendcounts = std::shared_ptr<std::vector<int>>(new std::vector<int>(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;
+}
- virtual void kernel(simgrid::xbt::ReplayAction& action) = 0;
+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
+ 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<std::vector<int>>(new std::vector<int>(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);
+}
- void* send_buffer(int size)
- {
- return smpi_get_tmp_sendbuffer(size);
- }
+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
+ 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<std::vector<int>>(new std::vector<int>(comm_size));
+ recvcounts = std::shared_ptr<std::vector<int>>(new std::vector<int>(comm_size));
+ senddisps = std::vector<int>(comm_size, 0);
+ recvdisps = std::vector<int>(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);
+}
- 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> {
-public:
- WaitAction(RequestStorage& storage) : ReplayAction("Wait", storage) {}
- void kernel(simgrid::xbt::ReplayAction& action) override
+ 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);
+ 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
if (is_wait_for_receive)
TRACE_smpi_recv(args.src, args.dst, args.tag);
}
-};
-class SendAction : public ReplayAction<SendRecvParser> {
-public:
- SendAction() = delete;
- explicit SendAction(std::string name, RequestStorage& storage) : ReplayAction(name, 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();
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());
}
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class RecvAction : public ReplayAction<SendRecvParser> {
-public:
- RecvAction() = delete;
- explicit RecvAction(std::string name, RequestStorage& storage) : ReplayAction(name, 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();
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);
TRACE_smpi_recv(src_traced, my_proc_id, args.tag);
}
}
-};
-class ComputeAction : public ReplayAction<ComputeParser> {
-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<WaitTestParser> {
-public:
- TestAction(RequestStorage& storage) : ReplayAction("Test", storage) {}
- void kernel(simgrid::xbt::ReplayAction& action) override
+ void TestAction::kernel(simgrid::xbt::ReplayAction& action)
{
- 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.
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);
}
}
-};
-class InitAction : public ReplayAction<ActionArgParser> {
-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
/* start a simulated timer */
smpi_process()->simulated_start();
- set_reqq_self(new std::vector<MPI_Request>);
}
-};
-class CommunicatorAction : public ReplayAction<ActionArgParser> {
-public:
- CommunicatorAction() : ReplayAction("Comm") {}
- void kernel(simgrid::xbt::ReplayAction& action) override { /* nothing to do */}
-};
+ void CommunicatorAction::kernel(simgrid::xbt::ReplayAction& action)
+ {
+ /* nothing to do */
+ }
-class WaitAllAction : public ReplayAction<ActionArgParser> {
-public:
- WaitAllAction(RequestStorage& storage) : ReplayAction("waitAll", storage) {}
- void kernel(simgrid::xbt::ReplayAction& action) override
+ void WaitAllAction::kernel(simgrid::xbt::ReplayAction& action)
{
- 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<std::pair</*sender*/int,/*recv*/int>> sender_receiver;
- for (const auto& req : (*get_reqq_self())) {
+ std::vector<MPI_Request> 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);
TRACE_smpi_comm_out(my_proc_id);
}
}
-};
-class BarrierAction : public ReplayAction<ActionArgParser> {
-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<BcastArgParser> {
-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(),
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class ReduceAction : public ReplayAction<ReduceArgParser> {
-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(),
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class AllReduceAction : public ReplayAction<AllReduceArgParser> {
-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), ""));
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class AllToAllAction : public ReplayAction<AllToAllArgParser> {
-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,
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class GatherAction : public ReplayAction<GatherArgParser> {
-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)));
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class GatherVAction : public ReplayAction<GatherVArgParser> {
-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();
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class ScatterAction : public ReplayAction<ScatterArgParser> {
-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,
TRACE_smpi_comm_out(my_proc_id);
}
-};
-
-class ScatterVAction : public ReplayAction<ScatterVArgParser> {
-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,
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class ReduceScatterAction : public ReplayAction<ReduceScatterArgParser> {
-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,
smpi_execute_flops(args.comp_size);
TRACE_smpi_comm_out(my_proc_id);
}
-};
-class AllToAllVAction : public ReplayAction<AllToAllVArgParser> {
-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(
TRACE_smpi_comm_out(my_proc_id);
}
-};
} // Replay Namespace
}} // namespace simgrid::smpi
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); });
/* 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){
