Updating to control the SHM allocation scheme under configure time options

configure.ac

@@ -331,8 +331,41 @@ case ${ac_PRECISION} in
      double)
        AC_DEFINE([GRID_DEFAULT_PRECISION_DOUBLE],[1],[GRID_DEFAULT_PRECISION is DOUBLE] )
      ;;
+     *)
+     AC_MSG_ERROR([${ac_PRECISION} unsupported --enable-precision option]);
+     ;;
 esac
 
+######################  Shared memory allocation technique under MPI3
+AC_ARG_ENABLE([shm],[AC_HELP_STRING([--enable-shm=shmget|shmopen|hugetlbfs],
+              [Select SHM allocation technique])],[ac_SHM=${enable_shm}],[ac_SHM=shmopen])
+
+case ${ac_SHM} in
+
+     shmget)
+     AC_DEFINE([GRID_MPI3_SHMGET],[1],[GRID_MPI3_SHMGET] )
+     ;;
+
+     shmopen)
+     AC_DEFINE([GRID_MPI3_SHMOPEN],[1],[GRID_MPI3_SHMOPEN] )
+     ;;
+
+     hugetlbfs)
+     AC_DEFINE([GRID_MPI3_SHMMMAP],[1],[GRID_MPI3_SHMMMAP] )
+     ;;
+
+     *)
+     AC_MSG_ERROR([${ac_SHM} unsupported --enable-shm option]);
+     ;;
+esac
+
+######################  Shared base path for SHMMMAP
+AC_ARG_ENABLE([shmpath],[AC_HELP_STRING([--enable-shmpath=path],
+              [Select SHM mmap base path for hugetlbfs])],
+	      [ac_SHMPATH=${enable_shmpath}],
+	      [ac_SHMPATH=/var/lib/hugetlbfs/pagesize-2MB/])
+AC_DEFINE_UNQUOTED([GRID_SHM_PATH],["$ac_SHMPATH"],[Path to a hugetlbfs filesystem for MMAPing])
+
 ############### communication type selection
 AC_ARG_ENABLE([comms],[AC_HELP_STRING([--enable-comms=none|mpi|mpi-auto|mpi3|mpi3-auto|shmem],
               [Select communications])],[ac_COMMS=${enable_comms}],[ac_COMMS=none])
@@ -482,6 +515,8 @@ compiler version            : ${ax_cv_gxx_version}
 SIMD                        : ${ac_SIMD}${SIMD_GEN_WIDTH_MSG}
 Threading                   : ${ac_openmp}
 Communications type         : ${comms_type}
+Shared memory allocator     : ${ac_SHM}
+Shared memory mmap path     : ${ac_SHMPATH}
 Default precision           : ${ac_PRECISION}
 Software FP16 conversion    : ${ac_SFW_FP16}
 RNG choice                  : ${ac_RNG}

lib/communicator/Communicator_mpi3.cc

@@ -198,7 +198,39 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
   ShmCommBuf = 0;
   ShmCommBufs.resize(ShmSize);
 
-#if 0
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // 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);
+    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_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 ) {       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++){
@@ -218,7 +250,7 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
 #endif
       void * ptr =  mmap(NULL,size, PROT_READ | PROT_WRITE, mmap_flag, 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);
 
 // Experiments; Experiments; Try to force numa domain on the shm segment if we have numaif.h
@@ -262,8 +294,11 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       ShmCommBufs[r] =ptr;
     }
   }
-
+#endif
-#else
+  ////////////////////////////////////////////////////////////////////////////////////////////
+  // SHMGET SHMAT and SHM_HUGETLB flag
+  ////////////////////////////////////////////////////////////////////////////////////////////
+#ifdef GRID_MPI3_SHMGET
   std::vector<int> shmids(ShmSize);
 
   if ( ShmRank == 0 ) {