Merge branch 'hotfix/dirac-ITT-fix1'

benchmarks/Benchmark_ITT.cc

@@ -232,9 +232,13 @@ public:
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
     std::cout<<GridLogMessage << "= Benchmarking a*x + y bandwidth"<<std::endl;
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
+    std::cout<<GridLogMessage << "  L  "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<< "\t\tGB/s / node"<<std::endl;
     std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
   
+    uint64_t NP;
+    uint64_t NN;
+
+
   uint64_t lmax=48;
 #define NLOOP (100*lmax*lmax*lmax*lmax/lat/lat/lat/lat)
 
@@ -245,6 +249,9 @@ public:
       int64_t vol= latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
       GridCartesian     Grid(latt_size,simd_layout,mpi_layout);
 
+      NP= Grid.RankCount();
+      NN =Grid.NodeCount();
+
       Vec rn ; random(sRNG,rn);
 
       LatticeVec z(&Grid); z=rn;
@@ -266,7 +273,8 @@ public:
       double flops=vol*Nvec*2;// mul,add
       double bytes=3.0*vol*Nvec*sizeof(Real);
       std::cout<<GridLogMessage<<std::setprecision(3) 
-	       << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.<<std::endl;
+	       << lat<<"\t\t"<<bytes<<"   \t\t"<<bytes/time<<"\t\t"<<flops/time<<"\t\t"<<(stop-start)/1000./1000.
+	       << "\t\t"<< bytes/time/NN <<std::endl;
 
     }
   };
@@ -387,6 +395,8 @@ public:
 	std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
 
 	int nwarm = 100;
+	uint64_t ncall = 1000;
+
 	double t0=usecond();
 	sFGrid->Barrier();
 	for(int i=0;i<nwarm;i++){
@@ -394,15 +404,8 @@ public:
 	}
 	sFGrid->Barrier();
 	double t1=usecond();
-	//	uint64_t ncall = (uint64_t) 2.5*1000.0*1000.0*nwarm/(t1-t0);
-	//	if (ncall < 500) ncall = 500;
-	uint64_t ncall = 500;
 
-	sFGrid->Broadcast(0,&ncall,sizeof(ncall));
-
-	//	std::cout << GridLogMessage << " Estimate " << ncall << " calls per second"<<std::endl;
 	sDw.ZeroCounters();
-
 	time_statistics timestat;
 	std::vector<double> t_time(ncall);
 	for(uint64_t i=0;i<ncall;i++){
@@ -436,12 +439,14 @@ public:
 	sDw.Report();
 
       }
+      double robust = mflops_worst/mflops_best;;
       std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
       std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Best  mflop/s        =   "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
       std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " sDeo Worst mflop/s        =   "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
-      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Performance Robustness   =   "<< mflops_worst/mflops_best <<std::endl;
+
+      std::cout<<GridLogMessage <<std::setprecision(3)<< L<<"^4 x "<<Ls<< " Performance Robustness   =   "<< robust <<std::endl;
       std::cout<<GridLogMessage <<fmt << std::endl;
-      std::cout<<GridLogMessage ;
+      std::cout<<GridLogMessage;
 
       for(int i=0;i<mflops_all.size();i++){
 	std::cout<<mflops_all[i]/NN<<" ; " ;
@@ -453,7 +458,7 @@ public:
     return mflops_best;
   }
 
-  static double DWF(int Ls,int L)
+  static double DWF(int Ls,int L, double & robust)
   {
     RealD mass=0.1;
     RealD M5  =1.8;
@@ -655,10 +660,11 @@ public:
 	assert((norm2(err)<1.0e-4));
 
       }
+      robust = mflops_worst/mflops_best;
       std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
       std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Best  mflop/s        =   "<< mflops_best << " ; " << mflops_best/NN<<" per node " <<std::endl;
       std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Deo Worst mflop/s        =   "<< mflops_worst<< " ; " << mflops_worst/NN<<" per node " <<std::endl;
-      std::cout<<GridLogMessage << L<<"^4 x "<<Ls<< " Performance Robustness   =   "<< mflops_worst/mflops_best <<std::endl;
+      std::cout<<GridLogMessage << std::fixed<<std::setprecision(3)<< L<<"^4 x "<<Ls<< " Performance Robustness   =   "<< robust  <<std::endl;
       std::cout<<GridLogMessage <<fmt << std::endl;
       std::cout<<GridLogMessage ;
 
@@ -692,26 +698,19 @@ int main (int argc, char ** argv)
   int do_wilson=1;
   int do_dwf   =1;
 
-  if ( do_memory ) {
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    std::cout<<GridLogMessage << " Memory benchmark " <<std::endl;
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    Benchmark::Memory();
-  }
-
-  if ( do_comms ) {
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    std::cout<<GridLogMessage << " Communications benchmark " <<std::endl;
-    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-    Benchmark::Comms();
-  }
-
   if ( do_su3 ) {
     // empty for now
   }
-
+#if 1
   int sel=2;
   std::vector<int> L_list({8,12,16,24});
+#else
+  int sel=1;
+  std::vector<int> L_list({8,12});
+#endif
+  int selm1=sel-1;
+  std::vector<double> robust_list;
+
   std::vector<double> wilson;
   std::vector<double> dwf4;
   std::vector<double> dwf5;
@@ -722,7 +721,8 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << " Wilson dslash 4D vectorised" <<std::endl;
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
     for(int l=0;l<L_list.size();l++){
-      wilson.push_back(Benchmark::DWF(1,L_list[l]));
+      double robust;
+      wilson.push_back(Benchmark::DWF(1,L_list[l],robust));
     }
   }
 
@@ -732,7 +732,10 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << " Domain wall dslash 4D vectorised" <<std::endl;
     std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
     for(int l=0;l<L_list.size();l++){
-      dwf4.push_back(Benchmark::DWF(Ls,L_list[l]));
+      double robust;
+      double result = Benchmark::DWF(Ls,L_list[l],robust) ;
+      dwf4.push_back(result);
+      robust_list.push_back(robust);
     }
   }
 
@@ -744,6 +747,10 @@ int main (int argc, char ** argv)
       dwf5.push_back(Benchmark::DWF5(Ls,L_list[l]));
     }
 
+  }
+
+  if ( do_dwf ) {
+
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
   std::cout<<GridLogMessage << " Summary table Ls="<<Ls <<std::endl;
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
@@ -752,11 +759,27 @@ int main (int argc, char ** argv)
     std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]<<" \t "<<dwf4[l]<<" \t "<<dwf5[l] <<std::endl;
   }
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+  }
 
+  int NN=NN_global;
+  if ( do_memory ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Memory benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::Memory();
+  }
+
+  if ( do_comms && (NN>1) ) {
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    std::cout<<GridLogMessage << " Communications benchmark " <<std::endl;
+    std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
+    Benchmark::Comms();
+  }
+
+  if ( do_dwf ) {
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
   std::cout<<GridLogMessage << " Per Node Summary table Ls="<<Ls <<std::endl;
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-  int NN=NN_global;
   std::cout<<GridLogMessage << " L \t\t Wilson\t\t DWF4  \t\t DWF5 " <<std::endl;
   for(int l=0;l<L_list.size();l++){
     std::cout<<GridLogMessage << L_list[l] <<" \t\t "<< wilson[l]/NN<<" \t "<<dwf4[l]/NN<<" \t "<<dwf5[l] /NN<<std::endl;
@@ -764,10 +787,12 @@ int main (int argc, char ** argv)
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
 
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
-  std::cout<<GridLogMessage << " Comparison point result: "  << dwf4[sel]/NN <<std::endl;
+  std::cout<<GridLogMessage << " Comparison point     result: "  << 0.5*(dwf4[sel]+dwf4[selm1])/NN << " Mflop/s per node"<<std::endl;
+  std::cout<<GridLogMessage << " Comparison point is 0.5*("<<dwf4[sel]/NN<<"+"<<dwf4[selm1]/NN << ") "<<std::endl;
+  std::cout<<std::setprecision(3);
+  std::cout<<GridLogMessage << " Comparison point robustness: "  << robust_list[sel] <<std::endl;
   std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
 
-
   }
 
 

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/allocator/AlignedAllocator.h

@@ -204,10 +204,12 @@ public:
 #endif
     size_type bytes = __n*sizeof(_Tp);
     uint8_t *cp = (uint8_t *)ptr;
+    if ( ptr ) { 
     // One touch per 4k page, static OMP loop to catch same loop order
 #pragma omp parallel for schedule(static)
-    for(size_type n=0;n<bytes;n+=4096){
-      cp[n]=0;
+      for(size_type n=0;n<bytes;n+=4096){
+	cp[n]=0;
+      }
     }
     return ptr;
   }

lib/communicator/Communicator_base.cc

@@ -37,7 +37,7 @@ namespace Grid {
 // Info that is setup once and indept of cartesian layout
 ///////////////////////////////////////////////////////////////
 void *              CartesianCommunicator::ShmCommBuf;
-uint64_t            CartesianCommunicator::MAX_MPI_SHM_BYTES   = 128*1024*1024; 
+uint64_t            CartesianCommunicator::MAX_MPI_SHM_BYTES   = 1024LL*1024LL*1024LL; 
 CartesianCommunicator::CommunicatorPolicy_t  
 CartesianCommunicator::CommunicatorPolicy= CartesianCommunicator::CommunicatorPolicyConcurrent;
 int CartesianCommunicator::nCommThreads = -1;
@@ -157,6 +157,9 @@ void CartesianCommunicator::ShmInitGeneric(void){
     perror("mmap failed ");
     exit(EXIT_FAILURE);  
   }
+#ifdef MADV_HUGEPAGE
+  if (!Hugepages ) madvise(ShmCommBuf,MAX_MPI_SHM_BYTES,MADV_HUGEPAGE);
+#endif
 #else 
   ShmBufStorageVector.resize(MAX_MPI_SHM_BYTES);
   ShmCommBuf=(void *)&ShmBufStorageVector[0];

lib/communicator/Communicator_mpi3.cc

@@ -198,7 +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++){
@@ -213,12 +252,15 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
       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);
 
-      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
@@ -240,7 +282,7 @@ void CartesianCommunicator::Init(int *argc, char ***argv) {
 	  if (ierr && (page==0)) perror("numa relocate command failed");
 	}
 #endif
-      ShmCommBufs[r] =ptr;
+	ShmCommBufs[r] =ptr;
       
     }
   }
@@ -262,25 +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;
+      key_t key   = IPC_PRIVATE;
       int flags = IPC_CREAT | SHM_R | SHM_W;
 #ifdef SHM_HUGETLB
-      flags|=SHM_HUGETLB;
+      if (Hugepages) flags|=SHM_HUGETLB;
 #endif
-      if ((shmids[r]= shmget(key,size, flags)) < 0) {
+      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);

lib/util/Init.cc

@@ -219,7 +219,8 @@ void Grid_init(int *argc,char ***argv)
     int MB;
     arg= GridCmdOptionPayload(*argv,*argv+*argc,"--shm");
     GridCmdOptionInt(arg,MB);
-    CartesianCommunicator::MAX_MPI_SHM_BYTES = MB*1024*1024;
+    uint64_t MB64 = MB;
+    CartesianCommunicator::MAX_MPI_SHM_BYTES = MB64*1024LL*1024LL;
   }
 
   if( GridCmdOptionExists(*argv,*argv+*argc,"--shm-hugepages") ){