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

Full reporting for benchmark; save robustness factor

Browse Source
This commit is contained in:
paboyle 2017-08-31 10:42:35 +01:00
parent 5b9267e88d
commit 7359df3501
4 changed files with 27 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
benchmarks/Benchmark_ITT.cc
View File

@ -439,10 +439,12 @@ 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;
@ -456,7 +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;
@ -658,10 +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 ;
@ -701,7 +704,11 @@ int main (int argc, char ** argv)
int sel=2;
std::vector<int> L_list({8,12,16,24});
//int sel=1;
// std::vector<int> L_list({8,12});
std::vector<double> robust_list;
std::vector<double> wilson;
std::vector<double> dwf4;
std::vector<double> dwf5;
@ -712,7 +719,8 @@ 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));
}
}
@ -722,7 +730,10 @@ 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);
}
}
@ -763,7 +774,6 @@ int main (int argc, char ** argv)
Benchmark::Comms();
}
if ( do_dwf ) {
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Per Node Summary table Ls="<<Ls <<std::endl;
@ -776,6 +786,8 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
std::cout<<GridLogMessage << " Comparison point result: " << dwf4[sel]/NN <<std::endl;
std::cout<<std::setprecision(3);
std::cout<<GridLogMessage << " Comparison point robustness: " << robust_list[sel]/NN <<std::endl;
std::cout<<GridLogMessage << "=================================================================================="<<std::endl;
}

2
lib/allocator/AlignedAllocator.h
View File

@ -204,11 +204,13 @@ 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;
}
}
return ptr;
}
void deallocate(pointer __p, size_type) {

2
lib/communicator/Communicator_base.cc
View File

@ -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;

3
lib/util/Init.cc
View File

@ -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") ){