MPI_File_read_shared, MPI_File_write_shared, MPI_File_read_ordered, MPI_File_write_or...

diff --git a/src/smpi/bindings/smpi_mpi.cpp b/src/smpi/bindings/smpi_mpi.cpp
index d1fb998..c11b325 100644 (file)
--- a/src/smpi/bindings/smpi_mpi.cpp
+++ b/src/smpi/bindings/smpi_mpi.cpp
@@ -377,14 +377,14 @@ UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_iwrite_all,(MPI_File fh, void *buf
 WRAPPED_PMPI_CALL(int, MPI_File_seek,(MPI_File fh, MPI_Offset offset, int whenace), (fh, offset, whenace))
 WRAPPED_PMPI_CALL(int, MPI_File_get_position,(MPI_File fh, MPI_Offset *offset), (fh, offset))
 UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_get_byte_offset,(MPI_File fh, MPI_Offset offset, MPI_Offset *disp), (fh, offset, disp))
-UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_read_shared,(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
-UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_write_shared,(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
+WRAPPED_PMPI_CALL(int, MPI_File_read_shared,(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
+WRAPPED_PMPI_CALL(int, MPI_File_write_shared,(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
 UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_iread_shared,(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Request *request), (fh, buf, count, datatype, request))
 UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_iwrite_shared,(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Request *request), (fh, buf, count, datatype, request))
-UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_read_ordered,(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
-UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_write_ordered,(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
-UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_seek_shared,(MPI_File fh, MPI_Offset offset, int whence), (fh, offset, whence))
-UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_get_position_shared,(MPI_File fh, MPI_Offset *offset), (fh, offset))
+WRAPPED_PMPI_CALL(int, MPI_File_read_ordered,(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
+WRAPPED_PMPI_CALL(int, MPI_File_write_ordered,(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status), (fh, buf, count, datatype, status))
+WRAPPED_PMPI_CALL(int, MPI_File_seek_shared,(MPI_File fh, MPI_Offset offset, int whence), (fh, offset, whence))
+WRAPPED_PMPI_CALL(int, MPI_File_get_position_shared,(MPI_File fh, MPI_Offset *offset), (fh, offset))
 UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_read_at_all_begin,(MPI_File fh, MPI_Offset offset, void *buf, int count, MPI_Datatype datatype), (fh, offset, buf, count, datatype))
 UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_read_at_all_end,(MPI_File fh, void *buf, MPI_Status *status), (fh, buf, status))
 UNIMPLEMENTED_WRAPPED_PMPI_CALL(int, MPI_File_write_at_all_begin,(MPI_File fh, MPI_Offset offset, void *buf, int count, MPI_Datatype datatype), (fh, offset, buf, count, datatype))

diff --git a/src/smpi/bindings/smpi_pmpi_file.cpp b/src/smpi/bindings/smpi_pmpi_file.cpp
index 1317eb7..ce31d25 100644 (file)
--- a/src/smpi/bindings/smpi_pmpi_file.cpp
+++ b/src/smpi/bindings/smpi_pmpi_file.cpp
@@ -61,6 +61,15 @@ int PMPI_File_seek(MPI_File fh, MPI_Offset offset, int whence){
   }\r
 }\r
 \r
+int PMPI_File_seek_shared(MPI_File fh, MPI_Offset offset, int whence){\r
+  CHECK_FILE(fh);\r
+  else {\r
+    smpi_bench_end();\r
+    int ret = fh->seek_shared(offset,whence);\r
+    smpi_bench_begin();\r
+    return ret;\r
+  }\r
+}\r
 \r
 int PMPI_File_get_position(MPI_File fh, MPI_Offset* offset){\r
   CHECK_FILE(fh);\r
@@ -74,6 +83,18 @@ int PMPI_File_get_position(MPI_File fh, MPI_Offset* offset){
   }\r
 }\r
 \r
+int PMPI_File_get_position_shared(MPI_File fh, MPI_Offset* offset){\r
+  CHECK_FILE(fh);\r
+  else if (offset==nullptr)\r
+    return MPI_ERR_DISP;\r
+  else {\r
+    smpi_bench_end();\r
+    int ret = fh->get_position_shared(offset);\r
+    smpi_bench_begin();\r
+    return ret;\r
+  }\r
+}\r
+\r
 int PMPI_File_read(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
   CHECK_FILE(fh);
   CHECK_BUFFER(buf, count);\r
@@ -93,6 +114,25 @@ int PMPI_File_read(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_
   }\r
 }\r
 \r
+int PMPI_File_read_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
+  CHECK_FILE(fh);\r
+  CHECK_BUFFER(buf, count);\r
+  CHECK_COUNT(count);\r
+  PASS_ZEROCOUNT(count);\r
+  CHECK_DATATYPE(datatype, count);\r
+  CHECK_STATUS(status);\r
+  CHECK_FLAGS(fh);\r
+  else {\r
+    smpi_bench_end();\r
+    int rank_traced = simgrid::s4u::this_actor::get_pid();\r
+    TRACE_smpi_comm_in(rank_traced, __func__, new simgrid::instr::CpuTIData("IO - read_shared", static_cast<double>(count*datatype->size())));\r
+    int ret = simgrid::smpi::File::read_shared(fh, buf, count, datatype, status);\r
+    TRACE_smpi_comm_out(rank_traced);\r
+    smpi_bench_begin();\r
+    return ret;\r
+  }\r
+}\r
+\r
 int PMPI_File_write(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
   CHECK_FILE(fh);
   CHECK_BUFFER(buf, count);\r
@@ -112,6 +152,25 @@ int PMPI_File_write(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI
   }\r
 }\r
 \r
+int PMPI_File_write_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
+  CHECK_FILE(fh);\r
+  CHECK_BUFFER(buf, count);\r
+  CHECK_COUNT(count);\r
+  PASS_ZEROCOUNT(count);\r
+  CHECK_DATATYPE(datatype, count);\r
+  CHECK_STATUS(status);\r
+  CHECK_FLAGS(fh);\r
+  else {\r
+    smpi_bench_end();\r
+    int rank_traced = simgrid::s4u::this_actor::get_pid();\r
+    TRACE_smpi_comm_in(rank_traced, __func__, new simgrid::instr::CpuTIData("IO - write_shared", static_cast<double>(count*datatype->size())));\r
+    int ret = simgrid::smpi::File::write_shared(fh, buf, count, datatype, status);\r
+    TRACE_smpi_comm_out(rank_traced);\r
+    smpi_bench_begin();\r
+    return ret;\r
+  }\r
+}\r
+\r
 int PMPI_File_read_all(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
   CHECK_FILE(fh);
   CHECK_BUFFER(buf, count);\r
@@ -130,6 +189,24 @@ int PMPI_File_read_all(MPI_File fh, void *buf, int count,MPI_Datatype datatype,
   }\r
 }\r
 \r
+int PMPI_File_read_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
+  CHECK_FILE(fh);\r
+  CHECK_BUFFER(buf, count);\r
+  CHECK_COUNT(count);\r
+  CHECK_DATATYPE(datatype, count);\r
+  CHECK_STATUS(status);\r
+  CHECK_FLAGS(fh);\r
+  else {\r
+    smpi_bench_end();\r
+    int rank_traced = simgrid::s4u::this_actor::get_pid();\r
+    TRACE_smpi_comm_in(rank_traced, __func__, new simgrid::instr::CpuTIData("IO - read_ordered", static_cast<double>(count*datatype->size())));\r
+    int ret = simgrid::smpi::File::read_ordered(fh, buf, count, datatype, status);\r
+    TRACE_smpi_comm_out(rank_traced);\r
+    smpi_bench_begin();\r
+    return ret;\r
+  }\r
+}\r
+\r
 int PMPI_File_write_all(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
   CHECK_FILE(fh);
   CHECK_BUFFER(buf, count);\r
@@ -148,6 +225,24 @@ int PMPI_File_write_all(MPI_File fh, void *buf, int count,MPI_Datatype datatype,
   }\r
 }\r
 \r
+int PMPI_File_write_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
+  CHECK_FILE(fh);\r
+  CHECK_BUFFER(buf, count);\r
+  CHECK_COUNT(count);\r
+  CHECK_DATATYPE(datatype, count);\r
+  CHECK_STATUS(status);\r
+  CHECK_FLAGS(fh);\r
+  else {\r
+    smpi_bench_end();\r
+    int rank_traced = simgrid::s4u::this_actor::get_pid();\r
+    TRACE_smpi_comm_in(rank_traced, __func__, new simgrid::instr::CpuTIData("IO - write_ordered", static_cast<double>(count*datatype->size())));\r
+    int ret = simgrid::smpi::File::write_ordered(fh, buf, count, datatype, status);\r
+    TRACE_smpi_comm_out(rank_traced);\r
+    smpi_bench_begin();\r
+    return ret;\r
+  }\r
+}\r
+\r
 int PMPI_File_read_at(MPI_File fh, MPI_Offset offset, void *buf, int count,MPI_Datatype datatype, MPI_Status *status){\r
   CHECK_FILE(fh);
   CHECK_BUFFER(buf, count);\r

diff --git a/src/smpi/include/smpi_file.hpp b/src/smpi/include/smpi_file.hpp
index 7c89f8b..42fd949 100644 (file)
--- a/src/smpi/include/smpi_file.hpp
+++ b/src/smpi/include/smpi_file.hpp
@@ -22,18 +22,28 @@ class File{
   simgrid::s4u::File* file_;\r
   MPI_Info info_;\r
   MPI_Offset shared_file_pointer_;\r
+  MPI_Win win_;\r
+  char* list_;\r
   public:\r
   File(MPI_Comm comm, char *filename, int amode, MPI_Info info);\r
   ~File();\r
   int size();\r
   int get_position(MPI_Offset* offset);\r
+  int get_position_shared(MPI_Offset* offset);\r
   int flags();\r
   int sync();\r
   int seek(MPI_Offset offset, int whence);\r
+  int seek_shared(MPI_Offset offset, int whence);\r
+  int lock();\r
+  int unlock();\r
   MPI_Info info();\r
   void set_info( MPI_Info info);\r
   static int read(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
+  static int read_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
+  static int read_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
   static int write(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
+  static int write_shared(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
+  static int write_ordered(MPI_File fh, void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
   template <int (*T)(MPI_File, void *, int, MPI_Datatype, MPI_Status *)> int op_all(void *buf, int count,MPI_Datatype datatype, MPI_Status *status);\r
   static int close(MPI_File *fh);\r
   static int del(char *filename, MPI_Info info);\r

diff --git a/src/smpi/mpi/smpi_file.cpp b/src/smpi/mpi/smpi_file.cpp
index 6006b4d..3aec0f2 100644 (file)
--- a/src/smpi/mpi/smpi_file.cpp
+++ b/src/smpi/mpi/smpi_file.cpp
@@ -8,18 +8,31 @@
 #include "smpi_coll.hpp"\r
 #include "smpi_datatype.hpp"\r
 #include "smpi_info.hpp"\r
+#include "smpi_win.hpp"\r
 #include "smpi_file.hpp"\r
+#include "smpi_status.hpp"\r
 #include "simgrid/plugins/file_system.h"\r
 \r
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_io, smpi, "Logging specific to SMPI (RMA operations)");\r
+#define FP_SIZE sizeof(MPI_Offset)\r
 \r
 \r
 namespace simgrid{\r
 namespace smpi{\r
-  File::File(MPI_Comm comm, char *filename, int amode, MPI_Info info): comm_(comm), flags_(amode), info_(info){\r
+\r
+  File::File(MPI_Comm comm, char *filename, int amode, MPI_Info info): comm_(comm), flags_(amode), info_(info), shared_file_pointer_(0) {\r
     file_= new simgrid::s4u::File(filename, nullptr);\r
+    list_=nullptr;\r
+    if (comm_->rank() == 0) {\r
+      int size= comm_->size() + FP_SIZE;\r
+      list_ = new char[size];\r
+      memset(list_, 0, size);\r
+      win_=new Win(list_, size, 1, MPI_INFO_NULL, comm_);\r
+    }else{\r
+      win_=new Win(list_, 0, 1, MPI_INFO_NULL, comm_);\r
+    }\r
   }\r
-  \r
+\r
   File::~File(){\r
     delete file_;\r
   }\r
@@ -45,6 +58,13 @@ namespace smpi{
     return MPI_SUCCESS;\r
   }\r
 \r
+  int File::get_position_shared(MPI_Offset* offset){\r
+    lock();\r
+    *offset=shared_file_pointer_;\r
+    unlock();\r
+    return MPI_SUCCESS;\r
+  }\r
+\r
   int File::seek(MPI_Offset offset, int whence){\r
     switch(whence){\r
       case(MPI_SEEK_SET):\r
@@ -64,7 +84,15 @@ namespace smpi{
     }\r
     return MPI_SUCCESS;\r
   }\r
-  \r
+\r
+  int File::seek_shared(MPI_Offset offset, int whence){\r
+    lock();\r
+    seek(offset,whence);\r
+    shared_file_pointer_=file_->tell();\r
+    unlock();\r
+    return MPI_SUCCESS;\r
+  }\r
+\r
   int File::read(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
     //get position first as we may be doing non contiguous reads and it will probably be updated badly\r
     MPI_Offset position = fh->file_->tell();\r
@@ -77,9 +105,39 @@ namespace smpi{
       fh->file_->seek(position+movesize, SEEK_SET);\r
     }\r
     XBT_DEBUG("Position after read in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());\r
+    status->count=count*datatype->size();\r
+    return MPI_SUCCESS;\r
+  }\r
+\r
+  int File::read_shared(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+    fh->lock();\r
+    fh->seek(fh->shared_file_pointer_,MPI_SEEK_SET);\r
+    read(fh, buf, count, datatype, status);\r
+    fh->shared_file_pointer_=fh->file_->tell();\r
+    fh->unlock();\r
     return MPI_SUCCESS;\r
   }\r
 \r
+  int File::read_ordered(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+    //0 needs to get the shared pointer value\r
+    if(fh->comm_->rank()==0){\r
+      fh->lock();\r
+      fh->unlock();\r
+    }else{\r
+      fh->shared_file_pointer_=count*datatype->size();\r
+    }\r
+    MPI_Offset result;\r
+    simgrid::smpi::Colls::scan(&(fh->shared_file_pointer_), &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);\r
+    fh->seek(result, MPI_SEEK_SET);\r
+    int ret = fh->op_all<simgrid::smpi::File::read>(buf, count, datatype, status);\r
+    if(fh->comm_->rank()==fh->comm_->size()-1){\r
+      fh->lock();\r
+      fh->unlock();\r
+    }\r
+    char c;\r
+    simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);\r
+    return ret;\r
+  }\r
 \r
   int File::write(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
     //get position first as we may be doing non contiguous reads and it will probably be updated badly\r
@@ -93,9 +151,40 @@ namespace smpi{
       fh->file_->seek(position+movesize, SEEK_SET);\r
     }\r
     XBT_DEBUG("Position after write in MPI_File %s : %llu",fh->file_->get_path(), fh->file_->tell());\r
+    status->count=count*datatype->size();\r
     return MPI_SUCCESS;\r
   }\r
 \r
+  int File::write_shared(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+    fh->lock();\r
+    fh->seek(fh->shared_file_pointer_,MPI_SEEK_SET);\r
+    write(fh, buf, count, datatype, status);\r
+    fh->shared_file_pointer_=fh->file_->tell();\r
+    fh->unlock();\r
+    return MPI_SUCCESS;\r
+  }\r
+\r
+  int File::write_ordered(MPI_File fh, void *buf, int count, MPI_Datatype datatype, MPI_Status *status){\r
+    //0 needs to get the shared pointer value\r
+    if(fh->comm_->rank()==0){\r
+      fh->lock();\r
+      fh->unlock();\r
+    }else{\r
+      fh->shared_file_pointer_=count*datatype->size();\r
+    }\r
+    MPI_Offset result;\r
+    simgrid::smpi::Colls::scan(&(fh->shared_file_pointer_), &result, 1, MPI_OFFSET, MPI_SUM, fh->comm_);\r
+    fh->seek(result, MPI_SEEK_SET);\r
+    int ret = fh->op_all<simgrid::smpi::File::write>(buf, count, datatype, status);\r
+    if(fh->comm_->rank()==fh->comm_->size()-1){\r
+      fh->lock();\r
+      fh->unlock();\r
+    }\r
+    char c;\r
+    simgrid::smpi::Colls::bcast(&c, 1, MPI_BYTE, fh->comm_->size()-1, fh->comm_);\r
+    return ret;\r
+  }\r
+\r
   int File::size(){\r
     return file_->size();\r
   }\r
@@ -108,6 +197,55 @@ namespace smpi{
     //no idea\r
     return simgrid::smpi::Colls::barrier(comm_);\r
   }\r
+\r
+  int File::lock()\r
+{\r
+  int rank = comm_->rank();\r
+  int size = comm_->size();\r
+  char waitlist[size];\r
+  char lock = 1;\r
+  int tag=444;\r
+  int i;\r
+  win_->lock(MPI_LOCK_EXCLUSIVE, 0, 0);\r
+  win_->put(&lock, 1, MPI_CHAR, 0, FP_SIZE+rank, 1, MPI_CHAR);\r
+  win_->get(waitlist, size, MPI_CHAR, 0, FP_SIZE, size, MPI_CHAR);\r
+  win_->get(&shared_file_pointer_ , 1 , MPI_OFFSET , 0 , 0, 1, MPI_OFFSET);\r
+  win_->unlock(0);\r
+  for (i = 0; i < size; i++) {\r
+    if (waitlist[i] == 1 && i != rank) {\r
+      // wait for the lock\r
+      MPI_Recv(&lock, 1, MPI_CHAR, MPI_ANY_SOURCE, tag, comm_, MPI_STATUS_IGNORE);\r
+      break;\r
+    }\r
+  }\r
+  return 0;\r
+}\r
+\r
+int File::unlock()\r
+{\r
+  int rank = comm_->rank();\r
+  int size = comm_->size();\r
+  char waitlist[size];\r
+  char lock = 0;\r
+  int tag=444;\r
+  int i, next;\r
+  win_->lock(MPI_LOCK_EXCLUSIVE, 0, 0);\r
+  win_->put(&lock, 1, MPI_CHAR, 0, FP_SIZE+rank, 1, MPI_CHAR);\r
+  win_->get(waitlist, size, MPI_CHAR, 0, FP_SIZE, size, MPI_CHAR);\r
+  shared_file_pointer_=file_->tell();\r
+  win_->put(&shared_file_pointer_, 1 , MPI_OFFSET , 0 , 0, 1, MPI_OFFSET);\r
+\r
+  win_->unlock(0);\r
+  next = (rank + 1 + size) % size;\r
+  for (i = 0; i < size; i++, next = (next + 1) % size) {\r
+    if (waitlist[next] == 1) {\r
+      MPI_Send(&lock, 1, MPI_CHAR, next, tag, comm_);\r
+      break;\r
+    }\r
+  }\r
+  return 0;\r
+}\r
+\r
 MPI_Info File::info(){\r
   if(info_== MPI_INFO_NULL)\r
     info_ = new Info();\r