portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-09-19 17:51:04 +01:00
Code Releases Activity

Merge branch 'develop' into feature/dwf-multirhs

Browse Source
This commit is contained in:
paboyle
2017-10-02 11:41:49 +01:00
parent ac740f73ce fddeb29d6b
commit 4f8b6f26b4
160 changed files with 23407 additions and 9004 deletions
Download Patch File Download Diff File Expand all files Collapse all files

143
lib/communicator/Communicator_mpi3.cc
View File

@@ -37,11 +37,12 @@ Author: Peter Boyle <paboyle@ph.ed.ac.uk>
#include <sys/ipc.h>
#include <sys/shm.h>
#include <sys/mman.h>
//#include <zlib.h>
#ifndef SHM_HUGETLB
#define SHM_HUGETLB 04000
#include <zlib.h>
#ifdef HAVE_NUMAIF_H
#include <numaif.h>
#endif
namespace Grid {
///////////////////////////////////////////////////////////////////////////////////////////////////
@@ -197,7 +198,46 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
ShmCommBuf = 0;
ShmCommBufs.resize(ShmSize);
#if 1
////////////////////////////////////////////////////////////////////////////////////////////
// Hugetlbf and others map filesystems as mappable huge pages
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMMMAP
char shm_name [NAME_MAX];
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
sprintf(shm_name,GRID_SHM_PATH "/Grid_mpi3_shm_%d_%d",GroupRank,r);
//sprintf(shm_name,"/var/lib/hugetlbfs/group/wheel/pagesize-2MB/" "Grid_mpi3_shm_%d_%d",GroupRank,r);
// printf("Opening file %s \n",shm_name);
int fd=open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd == -1) {
printf("open %s failed\n",shm_name);
perror("open hugetlbfs");
exit(0);
}
int mmap_flag = MAP_SHARED ;
#ifdef MAP_POPULATE
mmap_flag|=MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if ( Hugepages ) mmap_flag |= MAP_HUGETLB;
#endif
void *ptr = (void *) mmap(NULL, MAX_MPI_SHM_BYTES, PROT_READ | PROT_WRITE, mmap_flag,fd, 0);
if ( ptr == (void *)MAP_FAILED ) {
printf("mmap %s failed\n",shm_name);
perror("failed mmap"); assert(0);
}
assert(((uint64_t)ptr&0x3F)==0);
ShmCommBufs[r] =ptr;
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////
// POSIX SHMOPEN ; as far as I know Linux does not allow EXPLICIT HugePages with this case
// tmpfs (Larry Meadows says) does not support explicit huge page, and this is used for
// the posix shm virtual file system
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMOPEN
char shm_name [NAME_MAX];
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
@@ -210,11 +250,39 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
int fd=shm_open(shm_name,O_RDWR|O_CREAT,0666);
if ( fd < 0 ) { perror("failed shm_open"); assert(0); }
ftruncate(fd, size);
int mmap_flag = MAP_SHARED;
#ifdef MAP_POPULATE
mmap_flag |= MAP_POPULATE;
#endif
#ifdef MAP_HUGETLB
if (Hugepages) mmap_flag |= MAP_HUGETLB;
#endif
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, fd, 0);
void * ptr = mmap(NULL,size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
if ( ptr == MAP_FAILED ) { perror("failed mmap"); assert(0); }
if ( ptr == (void * )MAP_FAILED ) { perror("failed mmap"); assert(0); }
assert(((uint64_t)ptr&0x3F)==0);
ShmCommBufs[r] =ptr;
// Experiments; Experiments; Try to force numa domain on the shm segment if we have numaif.h
#if 0
//#ifdef HAVE_NUMAIF_H
int status;
int flags=MPOL_MF_MOVE;
#ifdef KNL
int nodes=1; // numa domain == MCDRAM
// Find out if in SNC2,SNC4 mode ?
#else
int nodes=r; // numa domain == MPI ID
#endif
unsigned long count=1;
for(uint64_t page=0;page<size;page+=4096){
void *pages = (void *) ( page + (uint64_t)ptr );
uint64_t *cow_it = (uint64_t *)pages; *cow_it = 1;
ierr= move_pages(0,count, &pages,&nodes,&status,flags);
if (ierr && (page==0)) perror("numa relocate command failed");
}
#endif
ShmCommBufs[r] =ptr;
}
}
@@ -236,21 +304,32 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
ShmCommBufs[r] =ptr;
}
}
#else
#endif
////////////////////////////////////////////////////////////////////////////////////////////
// SHMGET SHMAT and SHM_HUGETLB flag
////////////////////////////////////////////////////////////////////////////////////////////
#ifdef GRID_MPI3_SHMGET
std::vector<int> shmids(ShmSize);
if ( ShmRank == 0 ) {
for(int r=0;r<ShmSize;r++){
size_t size = CartesianCommunicator::MAX_MPI_SHM_BYTES;
key_t key = 0x4545 + r;
if ((shmids[r]= shmget(key,size, SHM_HUGETLB | IPC_CREAT | SHM_R | SHM_W)) < 0) {
key_t key = IPC_PRIVATE;
int flags = IPC_CREAT | SHM_R | SHM_W;
#ifdef SHM_HUGETLB
if (Hugepages) flags|=SHM_HUGETLB;
#endif
if ((shmids[r]= shmget(key,size, flags)) ==-1) {
int errsv = errno;
printf("Errno %d\n",errsv);
printf("key %d\n",key);
printf("size %lld\n",size);
printf("flags %d\n",flags);
perror("shmget");
exit(1);
} else {
printf("shmid: 0x%x\n", shmids[r]);
}
printf("shmid: 0x%x\n", shmids[r]);
}
}
MPI_Barrier(ShmComm);
@@ -384,8 +463,14 @@ CartesianCommunicator::CartesianCommunicator(const std::vector<int> &processors)
{
int ierr;
communicator=communicator_world;
_ndimension = processors.size();
communicator_halo.resize (2*_ndimension);
for(int i=0;i<_ndimension*2;i++){
MPI_Comm_dup(communicator,&communicator_halo[i]);
}
////////////////////////////////////////////////////////////////
// Assert power of two shm_size.
////////////////////////////////////////////////////////////////
@@ -608,13 +693,27 @@ void CartesianCommunicator::SendToRecvFromBegin(std::vector<CommsRequest_t> &lis
}
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes)
double CartesianCommunicator::StencilSendToRecvFrom( void *xmit,
int dest,
void *recv,
int from,
int bytes,int dir)
{
std::vector<CommsRequest_t> list;
double offbytes = StencilSendToRecvFromBegin(list,xmit,dest,recv,from,bytes,dir);
StencilSendToRecvFromComplete(list,dir);
return offbytes;
}
double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsRequest_t> &list,
void *xmit,
int dest,
void *recv,
int from,
int bytes,int dir)
{
assert(dir < communicator_halo.size());
MPI_Request xrq;
MPI_Request rrq;
@@ -633,26 +732,26 @@ double CartesianCommunicator::StencilSendToRecvFromBegin(std::vector<CommsReques
gfrom = MPI_UNDEFINED;
#endif
if ( gfrom ==MPI_UNDEFINED) {
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator,&rrq);
ierr=MPI_Irecv(recv, bytes, MPI_CHAR,from,from,communicator_halo[dir],&rrq);
assert(ierr==0);
list.push_back(rrq);
off_node_bytes+=bytes;
}
if ( gdest == MPI_UNDEFINED ) {
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator,&xrq);
ierr =MPI_Isend(xmit, bytes, MPI_CHAR,dest,_processor,communicator_halo[dir],&xrq);
assert(ierr==0);
list.push_back(xrq);
off_node_bytes+=bytes;
}
if ( CommunicatorPolicy == CommunicatorPolicySequential ) {
this->StencilSendToRecvFromComplete(list);
this->StencilSendToRecvFromComplete(list,dir);
}
return off_node_bytes;
}
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall)
void CartesianCommunicator::StencilSendToRecvFromComplete(std::vector<CommsRequest_t> &waitall,int dir)
{
SendToRecvFromComplete(waitall);
}