portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-04-10 22:20:45 +01:00
Code Releases Activity

Better timing tweaks to give sensible results on 24 threads on Edison dual ivybridge nodes.

Browse Source
This commit is contained in:
paboyle 2015-09-28 16:09:04 -07:00
parent 9f4f65cb46
commit af89c40462
2 changed files with 59 additions and 37 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
benchmarks/Benchmark_memory_asynch.cc
View File

@ -25,9 +25,10 @@ int main (int argc, char ** argv)
std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t"<<"bytes/thread"<<"\t\t\t"<<"GB/s"<<"\t\t\t"<<"GB/s per thread"<<std::endl; std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t"<<"bytes/thread"<<"\t\t\t"<<"GB/s"<<"\t\t\t"<<"GB/s per thread"<<std::endl;
std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl; std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
for(int lat=4;lat<=16536;lat*=2){ const int lmax = 16536*16;
for(int lat=4;lat<=lmax;lat*=2){
int Nloop=16536*1024*4/lat; int Nloop=lmax*128*4/lat;
std::vector<int> latt_size ({2*mpi_layout[0],2*mpi_layout[1],4*mpi_layout[2],lat*mpi_layout[3]}); std::vector<int> latt_size ({2*mpi_layout[0],2*mpi_layout[1],4*mpi_layout[2],lat*mpi_layout[3]});
@ -37,11 +38,15 @@ int main (int argc, char ** argv)
Vec tsum; tsum = zero; Vec tsum; tsum = zero;
GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice();
std::vector<double> stop(threads); std::vector<double> stop(threads);
Vector<Vec> sum(threads); Vector<Vec> sum(threads);
std::vector<LatticeVec> x(threads,&Grid); std::vector<LatticeVec> x(threads,&Grid);
for(int t=0;t<threads;t++){
random(pRNG,x[t]);
}
double start=usecond(); double start=usecond();
PARALLEL_FOR_LOOP PARALLEL_FOR_LOOP
@ -64,6 +69,9 @@ PARALLEL_FOR_LOOP
if ( stop[t]<min_stop ) min_stop=stop[t]; if ( stop[t]<min_stop ) min_stop=stop[t];
if ( stop[t]>max_stop ) max_stop=stop[t]; if ( stop[t]>max_stop ) max_stop=stop[t];
} }
double max_time = (max_stop-start)/Nloop*1000; double max_time = (max_stop-start)/Nloop*1000;
double min_time = (min_stop-start)/Nloop*1000; double min_time = (min_stop-start)/Nloop*1000;

82
benchmarks/Benchmark_memory_bandwidth.cc
View File

@ -10,144 +10,158 @@ int main (int argc, char ** argv)
const int Nvec=8; const int Nvec=8;
typedef Lattice< iVector< vReal,Nvec> > LatticeVec; typedef Lattice< iVector< vReal,Nvec> > LatticeVec;
typedef iVector<vReal,Nvec> Vec;
int Nloop=1000;
Vec rn = zero;
std::vector<int> simd_layout = GridDefaultSimd(Nd,vReal::Nsimd()); std::vector<int> simd_layout = GridDefaultSimd(Nd,vReal::Nsimd());
std::vector<int> mpi_layout = GridDefaultMpi(); std::vector<int> mpi_layout = GridDefaultMpi();
int threads = GridThread::GetThreads(); int threads = GridThread::GetThreads();
std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl; std::cout<<GridLogMessage << "Grid is setup to use "<<threads<<" threads"<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking fused AXPY bandwidth ; sizeof(Real) "<<sizeof(Real)<<std::endl; std::cout<<GridLogMessage << "= Benchmarking fused AXPY bandwidth ; sizeof(Real) "<<sizeof(Real)<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<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 << "----------------------------------------------------------"<<std::endl; std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
uint64_t lmax=44;
for(int lat=4;lat<=32;lat+=4){ #define NLOOP (100*lmax*lmax*lmax*lmax/vol)
for(int lat=4;lat<=lmax;lat+=4){
std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]}); std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3]; int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
GridCartesian Grid(latt_size,simd_layout,mpi_layout); GridCartesian Grid(latt_size,simd_layout,mpi_layout);
//GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice(); uint64_t Nloop=NLOOP;
LatticeVec z(&Grid); //random(pRNG,z); GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice();
LatticeVec x(&Grid); //random(pRNG,x);
LatticeVec y(&Grid); //random(pRNG,y); LatticeVec z(&Grid); random(pRNG,z);
LatticeVec x(&Grid); random(pRNG,x);
LatticeVec y(&Grid); random(pRNG,y);
double a=2.0; double a=2.0;
double start=usecond(); double start=usecond();
for(int i=0;i<Nloop;i++){ for(int i=0;i<Nloop;i++){
// inline void axpy(Lattice<vobj> &ret,double a,const Lattice<vobj> &lhs,const Lattice<vobj> &rhs){
axpy(z,a,x,y); axpy(z,a,x,y);
x._odata[0]=z._odata[0]; // serial loop dependence to prevent optimise
y._odata[4]=z._odata[4];
} }
double stop=usecond(); double stop=usecond();
double time = (stop-start)/Nloop*1000; double time = (stop-start)/Nloop*1000;
double flops=vol*Nvec*2;// mul,add double flops=vol*Nvec*2;// mul,add
double bytes=3*vol*Nvec*sizeof(Real); double bytes=3*vol*Nvec*sizeof(Real);
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t"<<flops/time<<std::endl; 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;
} }
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking a*x + y bandwidth"<<std::endl; std::cout<<GridLogMessage << "= Benchmarking a*x + y bandwidth"<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<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 << "----------------------------------------------------------"<<std::endl; std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
for(int lat=4;lat<=32;lat+=4){ for(int lat=4;lat<=lmax;lat+=4){
std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]}); std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3]; int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
GridCartesian Grid(latt_size,simd_layout,mpi_layout); GridCartesian Grid(latt_size,simd_layout,mpi_layout);
//GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice(); GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice();
LatticeVec z(&Grid); //random(pRNG,z); LatticeVec z(&Grid); random(pRNG,z);
LatticeVec x(&Grid); //random(pRNG,x); LatticeVec x(&Grid); random(pRNG,x);
LatticeVec y(&Grid); //random(pRNG,y); LatticeVec y(&Grid); random(pRNG,y);
double a=2.0; double a=2.0;
uint64_t Nloop=NLOOP;
double start=usecond(); double start=usecond();
for(int i=0;i<Nloop;i++){ for(int i=0;i<Nloop;i++){
z=a*x-y; z=a*x-y;
x._odata[0]=z._odata[0]; // force serial dependency to prevent optimise away
y._odata[4]=z._odata[4];
} }
double stop=usecond(); double stop=usecond();
double time = (stop-start)/Nloop*1000; double time = (stop-start)/Nloop*1000;
double flops=vol*Nvec*2;// mul,add double flops=vol*Nvec*2;// mul,add
double bytes=3*vol*Nvec*sizeof(Real); double bytes=3*vol*Nvec*sizeof(Real);
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t"<<flops/time<<std::endl; 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;
} }
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking SCALE bandwidth"<<std::endl; std::cout<<GridLogMessage << "= Benchmarking SCALE bandwidth"<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<std::endl; std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<"\t\t seconds"<<std::endl;
for(int lat=4;lat<=lmax;lat+=4){
for(int lat=4;lat<=32;lat+=4){
std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]}); std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3]; int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
uint64_t Nloop=NLOOP;
GridCartesian Grid(latt_size,simd_layout,mpi_layout); GridCartesian Grid(latt_size,simd_layout,mpi_layout);
//GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice(); GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice();
LatticeVec z(&Grid); //random(pRNG,z); LatticeVec z(&Grid); random(pRNG,z);
LatticeVec x(&Grid); //random(pRNG,x); LatticeVec x(&Grid); random(pRNG,x);
LatticeVec y(&Grid); //random(pRNG,y); LatticeVec y(&Grid); random(pRNG,y);
RealD a=2.0; RealD a=2.0;
double start=usecond(); double start=usecond();
for(int i=0;i<Nloop;i++){ for(int i=0;i<Nloop;i++){
z=a*x; z=a*x;
x._odata[0]=z._odata[0]*2.0;
} }
double stop=usecond(); double stop=usecond();
double time = (stop-start)/Nloop*1000; double time = (stop-start)/Nloop*1000;
double bytes=2*vol*Nvec*sizeof(Real); double bytes=2*vol*Nvec*sizeof(Real);
double flops=vol*Nvec*1;// mul double flops=vol*Nvec*1;// mul
std::cout<<GridLogMessage <<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t"<<flops/time<<std::endl; 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;
} }
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << "= Benchmarking READ bandwidth"<<std::endl; std::cout<<GridLogMessage << "= Benchmarking READ bandwidth"<<std::endl;
std::cout<<GridLogMessage << "===================================================================================================="<<std::endl; std::cout<<GridLogMessage << "===================================================================================================="<<std::endl;
std::cout<<GridLogMessage << " L "<<"\t\t"<<"bytes"<<"\t\t\t"<<"GB/s"<<"\t\t"<<"Gflop/s"<<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 << "----------------------------------------------------------"<<std::endl; std::cout<<GridLogMessage << "----------------------------------------------------------"<<std::endl;
for(int lat=4;lat<=32;lat+=4){ for(int lat=4;lat<=lmax;lat+=4){
std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]}); std::vector<int> latt_size ({lat*mpi_layout[0],lat*mpi_layout[1],lat*mpi_layout[2],lat*mpi_layout[3]});
int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3]; int vol = latt_size[0]*latt_size[1]*latt_size[2]*latt_size[3];
uint64_t Nloop=NLOOP;
GridCartesian Grid(latt_size,simd_layout,mpi_layout); GridCartesian Grid(latt_size,simd_layout,mpi_layout);
//GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice(); GridParallelRNG pRNG(&Grid); pRNG.SeedRandomDevice();
LatticeVec z(&Grid); random(pRNG,z);
LatticeVec z(&Grid); //random(pRNG,z); LatticeVec x(&Grid); random(pRNG,x);
LatticeVec x(&Grid); //random(pRNG,x); LatticeVec y(&Grid); random(pRNG,y);
LatticeVec y(&Grid); //random(pRNG,y);
RealD a=2.0; RealD a=2.0;
ComplexD nn; Real nn;
double start=usecond(); double start=usecond();
for(int i=0;i<Nloop;i++){ for(int i=0;i<Nloop;i++){
nn=norm2(x); nn=norm2(x);
vsplat(x._odata[0]._internal[0],nn);
} }
double stop=usecond(); double stop=usecond();
double time = (stop-start)/Nloop*1000; double time = (stop-start)/Nloop*1000;
double bytes=vol*Nvec*sizeof(Real); double bytes=vol*Nvec*sizeof(Real);
double flops=vol*Nvec*2;// mul,add double flops=vol*Nvec*2;// mul,add
std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t"<<flops/time<<std::endl; std::cout<<GridLogMessage<<std::setprecision(3) << lat<<"\t\t"<<bytes<<" \t\t"<<bytes/time<<"\t\t"<<flops/time<< "\t\t"<<(stop-start)/1000./1000.<< "\t\t " <<std::endl;
} }