portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2026-09-05 01:09:35 +01:00
Code Releases Activity

IO updates for AllToAllV aggregation of contiguous chunks.

Browse Source 
May finally clean up the poor MPI2 IO performance issue that has been persistent.
This commit is contained in:
Peter Boyle
2026-08-11 13:33:04 -04:00
parent 7647576863
commit 41f5a02204
7 changed files with 1019 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Grid/communicator/Communicator_base.h
+10
View File
@@ -238,6 +238,16 @@ public:
}
void AllToAll(int dim ,void *in,void *out,uint64_t words,uint64_t bytes);
void AllToAll(void *in,void *out,uint64_t words ,uint64_t bytes);
////////////////////////////////////////////////////////////////////////////
// Variable count all to all. Counts and displacements are in units of
// "bytes" sized words and are indexed by rank within this communicator.
// For exchanges that are a permutation but do not divide evenly between
// ranks; AllToAll above is the uniform count special case.
////////////////////////////////////////////////////////////////////////////
void AllToAllV(void *in ,const std::vector<int> &sendcounts,const std::vector<int> &senddispls,
void *out,const std::vector<int> &recvcounts,const std::vector<int> &recvdispls,
uint64_t bytes);
template<class obj> void Broadcast(int root,obj &data)
{
Grid/communicator/Communicator_mpi3.cc
+20
View File
@@ -945,5 +945,25 @@ void CartesianCommunicator::AllToAll(void *in,void *out,uint64_t words,uint64_t
MPI_Alltoall(in,iwords,object,out,iwords,object,communicator);
MPI_Type_free(&object);
}
void CartesianCommunicator::AllToAllV(void *in ,const std::vector<int> &sendcounts,const std::vector<int> &senddispls,
void *out,const std::vector<int> &recvcounts,const std::vector<int> &recvdispls,
uint64_t bytes)
{
FlightRecorder::StepLog("AllToAllV");
GRID_ASSERT(sendcounts.size()==(size_t)_Nprocessors);
GRID_ASSERT(senddispls.size()==(size_t)_Nprocessors);
GRID_ASSERT(recvcounts.size()==(size_t)_Nprocessors);
GRID_ASSERT(recvdispls.size()==(size_t)_Nprocessors);
// MPI counts are "int"; the caller sizes the word to keep them in range
int ibytes = bytes;
GRID_ASSERT(bytes == (uint64_t)ibytes);
MPI_Datatype object;
MPI_Type_contiguous(ibytes,MPI_BYTE,&object);
MPI_Type_commit(&object);
int ierr = MPI_Alltoallv(in ,(int *)&sendcounts[0],(int *)&senddispls[0],object,
out,(int *)&recvcounts[0],(int *)&recvdispls[0],object,communicator);
GRID_ASSERT(ierr==0);
MPI_Type_free(&object);
}
NAMESPACE_END(Grid);
Grid/communicator/Communicator_none.cc
+11
View File
@@ -113,6 +113,17 @@ void CartesianCommunicator::AllToAll(void *in,void *out,uint64_t words,uint64_t
{
bcopy(in,out,bytes*words);
}
void CartesianCommunicator::AllToAllV(void *in ,const std::vector<int> &sendcounts,const std::vector<int> &senddispls,
void *out,const std::vector<int> &recvcounts,const std::vector<int> &recvdispls,
uint64_t bytes)
{
// Single rank: the exchange degenerates to a copy of our own segment
GRID_ASSERT(sendcounts.size()==1);
GRID_ASSERT(recvcounts.size()==1);
GRID_ASSERT(sendcounts[0]==recvcounts[0]);
bcopy((char *)in +(uint64_t)senddispls[0]*bytes,
(char *)out+(uint64_t)recvdispls[0]*bytes,bytes*(uint64_t)sendcounts[0]);
}
int CartesianCommunicator::RankWorld(void){return 0;}
void CartesianCommunicator::Barrier(void){}