Basic level NUMA support (build and runtime availability check).

Author: theurich

build/common.mk

@@ -1811,6 +1811,38 @@ ESMF_F90LINKRPATHSTHIRD   += $(ESMF_F90RPATHPREFIX)$(ESMF_BABELTRACE_LIBPATH)
 endif
 endif
 
+#-------------------------------------------------------------------------------
+# NUMA
+#-------------------------------------------------------------------------------
+ifeq ($(ESMF_NUMA),ON)
+ESMF_NUMA = standard
+endif
+ifeq ($(ESMF_NUMA),standard)
+ifneq ($(origin ESMF_NUMA_LIBS), environment)
+ESMF_NUMA_LIBS = -lnuma
+endif
+endif
+
+ifdef ESMF_NUMA
+ESMF_CPPFLAGS                += -DESMF_NUMA=1
+ifdef ESMF_NUMA_INCLUDE
+ESMF_CXXCOMPILEPATHSTHIRD    += -I$(ESMF_NUMA_INCLUDE)
+ESMF_F90COMPILEPATHSTHIRD    += -I$(ESMF_NUMA_INCLUDE)
+endif
+ifdef ESMF_NUMA_LIBS
+ESMF_CXXLINKLIBS          += $(ESMF_NUMA_LIBS)
+ESMF_CXXLINKRPATHSTHIRD   += $(addprefix $(ESMF_CXXRPATHPREFIX),$(subst -L,,$(filter -L%,$(ESMF_NUMA_LIBS))))
+ESMF_F90LINKLIBS          += $(ESMF_NUMA_LIBS)
+ESMF_F90LINKRPATHSTHIRD   += $(addprefix $(ESMF_F90RPATHPREFIX),$(subst -L,,$(filter -L%,$(ESMF_NUMA_LIBS))))
+endif
+ifdef ESMF_NUMA_LIBPATH
+ESMF_CXXLINKPATHSTHIRD    += -L$(ESMF_NUMA_LIBPATH)
+ESMF_F90LINKPATHSTHIRD    += -L$(ESMF_NUMA_LIBPATH)
+ESMF_CXXLINKRPATHSTHIRD   += $(ESMF_CXXRPATHPREFIX)$(ESMF_NUMA_LIBPATH)
+ESMF_F90LINKRPATHSTHIRD   += $(ESMF_F90RPATHPREFIX)$(ESMF_NUMA_LIBPATH)
+endif
+endif
+
 #-------------------------------------------------------------------------------
 # NVML
 #-------------------------------------------------------------------------------

makefile

@@ -304,6 +304,18 @@ endif
 	    echo "ESMF_BABELTRACE_LIBPATH: $(ESMF_BABELTRACE_LIBPATH)" ; \
 	  fi; \
 	 fi
+	-@if [ -n "$(ESMF_NUMA)" ] ; then \
+	  echo "ESMF_NUMA:               $(ESMF_NUMA)" ; \
+	  if [ -n "$(ESMF_NUMA_INCLUDE)" ] ; then \
+	    echo "ESMF_NUMA_INCLUDE:       $(ESMF_NUMA_INCLUDE)" ; \
+	  fi; \
+	  if [ -n "$(ESMF_NUMA_LIBS)" ] ; then \
+	    echo "ESMF_NUMA_LIBS:          $(ESMF_NUMA_LIBS)" ; \
+	  fi; \
+	  if [ -n "$(ESMF_NUMA_LIBPATH)" ] ; then \
+	    echo "ESMF_NUMA_LIBPATH:       $(ESMF_NUMA_LIBPATH)" ; \
+	  fi; \
+	 fi
 	-@if [ -n "$(ESMF_NVML)" ] ; then \
 	  echo "ESMF_NVML:               $(ESMF_NVML)" ; \
 	  if [ -n "$(ESMF_NVML_INCLUDE)" ] ; then \
@@ -709,6 +721,18 @@ endif
 	    echo "# ESMF_BABELTRACE_LIBPATH:$(ESMF_BABELTRACE_LIBPATH)" >> $(MKINFO) ; \
 	  fi; \
 	 fi
+	-@if [ -n "$(ESMF_NUMA)" ] ; then \
+	  echo "# ESMF_NUMA:              $(ESMF_NUMA)" >> $(MKINFO) ; \
+	  if [ -n "$(ESMF_NUMA_INCLUDE)" ] ; then \
+	    echo "# ESMF_NUMA_INCLUDE:      $(ESMF_NUMA_INCLUDE)" >> $(MKINFO) ; \
+	  fi; \
+	  if [ -n "$(ESMF_NUMA_LIBS)" ] ; then \
+	    echo "# ESMF_NUMA_LIBS:         $(ESMF_NUMA_LIBS)" >> $(MKINFO) ; \
+	  fi; \
+	  if [ -n "$(ESMF_NUMA_LIBPATH)" ] ; then \
+	    echo "# ESMF_NUMA_LIBPATH:      $(ESMF_NUMA_LIBPATH)" >> $(MKINFO) ; \
+	  fi; \
+	 fi
 	-@if [ -n "$(ESMF_NVML)" ] ; then \
 	  echo "# ESMF_NVML:              $(ESMF_NVML)" >> $(MKINFO) ; \
 	  if [ -n "$(ESMF_NVML_INCLUDE)" ] ; then \

src/Infrastructure/VM/include/ESMCI_VMKernel.h

@@ -313,21 +313,26 @@ class VMK{
     int *ncpet;     // number of cores this pet references
     int *nadevs;//TODO: to be removed // number of accelerator devices accessible from this pet
     int **cid;      // core id of the cores this pet references
+    // SSI
     int ssiCount;   // number of single system images in this VMK
     int ssiMinPetCount;   // minimum PETs on a single system image
     int ssiMaxPetCount;   // maximum PETs on a single system image
     int ssiLocalPetCount; // number of PETs on the same SSI as localPet (incl.)
     int ssiLocalPet;      // id of local PET in the local SSI
     int *ssiLocalPetList; // PETs that are on the same SSI as localPet (incl.)
+    // SSI DEV
     int devCount;   // number of devices associated with this VMK all SSI
     int ssiLocalDevCount;// number of devices associated with this VMK on local SSI
     int *ssiLocalDevList;// list of SSI-local device indices associated with this VMK
                     // Use this index to make local association calls (e.g. via
                     // acc_set_device_num() or omp_set_default_device()), and
                     // to look up global device index in ssidevs array.
+    // SSI NUMA
+    int ssiLocalNumaCount; // number of NUMA modes on the same SSI as localPet (incl.)
+    int *ssiLocalNumaList; // NUMA nodes
     // general information about this VMK
-    int mpionly;    // 0: there is multi-threading, 1: MPI-only
-    bool threadsflag; // threaded or none-threaded VM
+    bool mpionly;         // false: there is multi-threading, true: MPI-only
+    bool threadsflag;     // threaded or none-threaded VM
     // MPI Communicator handles
     MPI_Comm mpi_c;     // communicator across the entire VM
     MPI_Comm mpi_c_ssi; // communicator holding PETs on the same SSI
@@ -399,6 +404,10 @@ class VMK{
     static int argc_mpich;
     static char *argv_mpich_store[100];
     static char **argv_mpich;
+    // NVML support
+    static bool nvmlEnabled;     // NVML support enabled or disabled
+    // NUMA support
+    static bool numaEnabled;     // NUMA support enabled or disabled
 
   // methods
   private:
@@ -515,6 +524,12 @@ class VMK{
       return false;
 #endif
     }
+    static bool isNvmlEnabled(){
+      return nvmlEnabled;
+    }
+    static bool isNumaEnabled(){
+      return numaEnabled;
+    }
 
 #define XSTR(X) STR(X)
 #define STR(X) #X

src/Infrastructure/VM/src/ESMCI_VMKernel.C

@@ -94,6 +94,12 @@ using namespace std;
 #include <openacc.h>
 #endif
 
+// NUMA support
+#ifdef ESMF_NUMA
+#include <numa.h>
+#endif
+
+// NVML support
 #ifdef ESMF_NVML
 #include <nvml.h>
 #endif
@@ -157,6 +163,10 @@ char **VMK::argv = &(argv_store[0]);
 int VMK::argc_mpich;
 char *VMK::argv_mpich_store[100];
 char **VMK::argv_mpich = &(argv_mpich_store[0]);
+// NUMA support
+bool VMK::nvmlEnabled;
+// NUMA support
+bool VMK::numaEnabled;
 
 } // namespace ESMCI
 
@@ -447,12 +457,12 @@ void VMK::init(MPI_Comm mpiCommunicator, bool globalResourceControl){
   mypthid=0;
 #endif
 #ifdef ESMF_MPIUNI
-  mpionly=0;          // this way the commtype will be checked in comm calls
+  mpionly=false;        // this way the commtype will be checked in comm calls
 #else
   if (npets==1)
-    mpionly=0;          // this way the commtype will be checked in comm calls
+    mpionly=false;      // this way the commtype will be checked in comm calls
   else
-    mpionly=1;          // normally the default VM can only be MPI-only
+    mpionly=true;       // normally the default VM can only be MPI-only
 #endif
   // no threading in default global VM
   threadsflag = false;
@@ -472,6 +482,19 @@ void VMK::init(MPI_Comm mpiCommunicator, bool globalResourceControl){
   MPI_Comm_rank(mpi_c_ssi, &color);
   if (color>0) color = MPI_UNDEFINED; // only root PETs on each SSI
   MPI_Comm_split(mpi_c, color, 0, &mpi_c_ssi_roots);
+#endif
+  // check whether NVML support is enabled
+#ifdef ESMF_NVML
+  nvmlEnabled = true;
+#else
+  nvmlEnabled = false;
+#endif
+  // check whether NUMA support is enabled
+#ifdef ESMF_NUMA
+  numaEnabled = true;
+  if (numa_available()<0) numaEnabled = false;  // NUMA not available at runtime
+#else
+  numaEnabled = false;
 #endif
   // initialize the shared memory variables
   pth_finish_count = NULL;
@@ -1101,15 +1124,15 @@ void VMK::construct(void *ssarg){
   }
 #ifdef ESMF_MPIUNI
   // don't set mpionly flag so that comm call check for commtype
-  mpionly=0;
+  mpionly=false;
 #else
   if (npets==1)
-    mpionly=0;
+    mpionly=false;
   else{
     // determine whether we are dealing with an MPI-only VMK
-    mpionly=1;  // assume this is MPI-only VMK until found otherwise
+    mpionly=true;  // assume this is MPI-only VMK until found otherwise
     for (int i=0; i<npets; i++)
-      if (tid[i]>0) mpionly=0;    // found multi-threading PET
+      if (tid[i]>0) mpionly=false;    // found multi-threading PET
   }
 #endif
   threadsflag = sarg->threadsflag;
@@ -3096,6 +3119,12 @@ void VMK::logSystem(std::string prefix, ESMC_LogMsgType_Flag msgType){
   msg << prefix << "isSsiSharedMemoryEnabled=" << isSsiSharedMemoryEnabled();
   ESMC_LogDefault.Write(msg.str(), msgType);
   msg.str("");  // clear
+  msg << prefix << "isNvmlEnabled=" << isNvmlEnabled();
+  ESMC_LogDefault.Write(msg.str(), msgType);
+  msg.str("");  // clear
+  msg << prefix << "isNumaEnabled=" << isNumaEnabled();
+  ESMC_LogDefault.Write(msg.str(), msgType);
+  msg.str("");  // clear
   msg << prefix << "ssiCount=" << nssiid << " peCount=" << ncores;
   ESMC_LogDefault.Write(msg.str(), msgType);
   for (int i=0; i<ncores; i++){