portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-09 23:45:36 +00:00
Code Releases Activity

Priintinig and device memory size detection

Browse Source
This commit is contained in:
Peter Boyle 2020-06-04 14:58:03 -04:00
parent 0d95805cde
commit f39c2a240b
4 changed files with 19 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
Grid/allocator/MemoryManager.h
View File

@ -101,9 +101,6 @@ private:
static void *CpuAllocate(size_t bytes);
static void CpuFree (void *ptr,size_t bytes);
private:
////////////////////////////////////////////////////////
// Footprint tracking
////////////////////////////////////////////////////////
@ -115,6 +112,7 @@ private:
static uint64_t HostToDeviceXfer;
static uint64_t DeviceToHostXfer;
private:
#ifndef GRID_UVM
//////////////////////////////////////////////////////////////////////
// Data tables for ViewCache

2
Grid/allocator/MemoryManagerCache.cc
View File

@ -439,7 +439,7 @@ void MemoryManager::Print(void)
std::cout << GridLogDebug << DeviceToHostXfer << " transfers from device " << std::endl;
std::cout << GridLogDebug << HostToDeviceBytes<< " bytes transfered to device " << std::endl;
std::cout << GridLogDebug << DeviceToHostBytes<< " bytes transfered from device " << std::endl;
std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << std::endl;
std::cout << GridLogDebug << AccViewTable.size()<< " vectors " << LRU.size()<<" evictable"<< std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;
std::cout << GridLogDebug << "CpuAddr\t\tAccAddr\t\tState\t\tcpuLock\taccLock\tLRU_valid "<<std::endl;
std::cout << GridLogDebug << "--------------------------------------------" << std::endl;

5
Grid/threads/Accelerator.cc
View File

@ -31,6 +31,7 @@ void acceleratorInit(void)
if ((localRankStr = getenv(ENV_RANK_OMPI )) != NULL) { world_rank = atoi(localRankStr);}
if ((localRankStr = getenv(ENV_RANK_MVAPICH)) != NULL) { world_rank = atoi(localRankStr);}
size_t totalDeviceMem=0;
for (int i = 0; i < nDevices; i++) {
#define GPU_PROP_FMT(canMapHostMemory,FMT) printf("AcceleratorCudaInit: " #canMapHostMemory ": " FMT" \n",prop.canMapHostMemory);
@ -45,14 +46,18 @@ void acceleratorInit(void)
printf("AcceleratorCudaInit: ========================\n");
printf("AcceleratorCudaInit: Device identifier: %s\n", prop.name);
GPU_PROP_FMT(totalGlobalMem,"%lld");
GPU_PROP(managedMemory);
GPU_PROP(isMultiGpuBoard);
GPU_PROP(warpSize);
totalDeviceMem = prop.totalGlobalMem;
// GPU_PROP(unifiedAddressing);
// GPU_PROP(l2CacheSize);
// GPU_PROP(singleToDoublePrecisionPerfRatio);
}
}
MemoryManager::DeviceMaxBytes = (8*totalDeviceMem)/10; // Assume 80% ours
#ifdef GRID_IBM_SUMMIT
// IBM Jsrun makes cuda Device numbering screwy and not match rank
if ( world_rank == 0 ) printf("AcceleratorCudaInit: IBM Summit or similar - NOT setting device to node rank\n");

12
Grid/util/Init.cc
View File

@ -296,6 +296,14 @@ void Grid_init(int *argc,char ***argv)
GlobalSharedMemory::MAX_MPI_SHM_BYTES = MB64*1024LL*1024LL;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--device-mem") ){
int GB;
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--device-mem");
GridCmdOptionInt(arg,GB);
uint64_t GB64 = GB;
MemoryManager::DeviceMaxBytes = GB64*1024LL*1024LL*1024LL;
}
if( GridCmdOptionExists(*argv,*argv+*argc,"--hypercube") ){
int enable;
arg= GridCmdOptionPayload(*argv,*argv+*argc,"--hypercube");
@ -355,6 +363,10 @@ void Grid_init(int *argc,char ***argv)
std::cout << GridLogMessage << "Mapped stencil comms buffers as MAP_HUGETLB "<<std::endl;
}
#ifndef GRID_UVM
std::cout << GridLogMessage << "MemoryManager Cache "<< MemoryManager::DeviceMaxBytes <<" bytes "<<std::endl;
#endif
if( GridCmdOptionExists(*argv,*argv+*argc,"--debug-mem") ){
MemoryProfiler::debug = true;
MemoryProfiler::stats = &dbgMemStats;