Fixing perf measures in TestCommD3Q19
This commit is contained in:
@@ -481,7 +481,7 @@ int main(int argc, char **argv)
|
|||||||
// cout << "CPU time: " << (stoptime - starttime) << " seconds" << endl;
|
// cout << "CPU time: " << (stoptime - starttime) << " seconds" << endl;
|
||||||
cputime = stoptime - starttime;
|
cputime = stoptime - starttime;
|
||||||
// cout << "Lattice update rate: "<< double(Nx*Ny*Nz*timestep)/cputime/1000000 << " MLUPS" << endl;
|
// cout << "Lattice update rate: "<< double(Nx*Ny*Nz*timestep)/cputime/1000000 << " MLUPS" << endl;
|
||||||
double MLUPS = double(Nx*Ny*Nz*timestep)/cputime/1000000;
|
double MLUPS = double(Np)*double(timestep)/cputime*1e-6;
|
||||||
if (rank==0) printf("********************************************************\n");
|
if (rank==0) printf("********************************************************\n");
|
||||||
if (rank==0) printf("CPU time = %f \n", cputime);
|
if (rank==0) printf("CPU time = %f \n", cputime);
|
||||||
if (rank==0) printf("Lattice update rate (per process)= %f MLUPS \n", MLUPS);
|
if (rank==0) printf("Lattice update rate (per process)= %f MLUPS \n", MLUPS);
|
||||||
@@ -491,9 +491,9 @@ int main(int argc, char **argv)
|
|||||||
|
|
||||||
// Number of memory references from the swap algorithm (per timestep)
|
// Number of memory references from the swap algorithm (per timestep)
|
||||||
// 18 reads and 18 writes for each lattice site
|
// 18 reads and 18 writes for each lattice site
|
||||||
double MemoryRefs = (Nx-2)*(Ny-2)*(Nz-2)*36;
|
double MemoryRefs = double(Np)*36;
|
||||||
// number of memory references for the swap algorithm - GigaBytes / second
|
// number of memory references for the swap algorithm - GigaBytes / second
|
||||||
if (rank==0) printf("DRAM bandwidth (per process)= %f GB/sec \n",MemoryRefs*8*timestep/1e9);
|
if (rank==0) printf("DRAM bandwidth (per process)= %f GB/sec \n",MemoryRefs*8*double(timestep)*1e-9);
|
||||||
// Report bandwidth in Gigabits per second
|
// Report bandwidth in Gigabits per second
|
||||||
// communication bandwidth includes both send and recieve
|
// communication bandwidth includes both send and recieve
|
||||||
if (rank==0) printf("Communication bandwidth (per process)= %f Gbit/sec \n",ScaLBL_Comm.CommunicationCount*64*timestep/1e9);
|
if (rank==0) printf("Communication bandwidth (per process)= %f Gbit/sec \n",ScaLBL_Comm.CommunicationCount*64*timestep/1e9);
|
||||||
|
|||||||
@@ -645,7 +645,7 @@ int main(int argc, char **argv)
|
|||||||
// Compute the walltime per timestep
|
// Compute the walltime per timestep
|
||||||
cputime = (stoptime - starttime)/timestep;
|
cputime = (stoptime - starttime)/timestep;
|
||||||
// Performance obtained from each node
|
// Performance obtained from each node
|
||||||
double MLUPS = double(Nx*Ny*Nz)/cputime/1000000;
|
double MLUPS = double(Np)/cputime/1000000;
|
||||||
|
|
||||||
if (rank==0) printf("********************************************************\n");
|
if (rank==0) printf("********************************************************\n");
|
||||||
if (rank==0) printf("CPU time = %f \n", cputime);
|
if (rank==0) printf("CPU time = %f \n", cputime);
|
||||||
@@ -655,16 +655,6 @@ int main(int argc, char **argv)
|
|||||||
if (rank==0) printf("********************************************************\n");
|
if (rank==0) printf("********************************************************\n");
|
||||||
|
|
||||||
// ************************************************************************
|
// ************************************************************************
|
||||||
double *PHASE;
|
|
||||||
int SIZE=Nx*Ny*Nz*sizeof(double);
|
|
||||||
PHASE= new double [Nx*Ny*Nz];
|
|
||||||
ScaLBL_CopyToHost(&PHASE[0],&Phi[0],SIZE);
|
|
||||||
|
|
||||||
FILE *OUTFILE;
|
|
||||||
sprintf(LocalRankFilename,"Phase.%05i.raw",rank);
|
|
||||||
OUTFILE = fopen(LocalRankFilename,"wb");
|
|
||||||
fwrite(PHASE,8,N,OUTFILE);
|
|
||||||
fclose(OUTFILE);
|
|
||||||
|
|
||||||
PROFILE_STOP("Main");
|
PROFILE_STOP("Main");
|
||||||
PROFILE_SAVE("lbpm_color_simulator",1);
|
PROFILE_SAVE("lbpm_color_simulator",1);
|
||||||
|
|||||||
@@ -424,7 +424,7 @@ int main(int argc, char **argv)
|
|||||||
// Compute the walltime per timestep
|
// Compute the walltime per timestep
|
||||||
cputime = (stoptime - starttime)/timestep;
|
cputime = (stoptime - starttime)/timestep;
|
||||||
// Performance obtained from each node
|
// Performance obtained from each node
|
||||||
double MLUPS = double(Nx*Ny*Nz)/cputime/1000000;
|
double MLUPS = double(Np)/cputime/1000000;
|
||||||
|
|
||||||
if (rank==0) printf("********************************************************\n");
|
if (rank==0) printf("********************************************************\n");
|
||||||
if (rank==0) printf("CPU time = %f \n", cputime);
|
if (rank==0) printf("CPU time = %f \n", cputime);
|
||||||
|
|||||||
Reference in New Issue
Block a user