[petsc-users] Memory usage during matrix factorization

[Barry Smith]

   These return what the operating system reports is being used by the process so it includes any external packages. It is for a single process if you want the value over a set of processes then use MP_Allreduce() to sum them up.

   Barry

   Here is the code: it is only as reliable as the OS is at reporting the values.


#if defined(PETSC_USE_PROCFS_FOR_SIZE)

  sprintf(proc,"/proc/%d",(int)getpid());
  if ((fd = open(proc,O_RDONLY)) == -1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_FILE_OPEN,"Unable to access system file %s to get memory usage data",file);
  if (ioctl(fd,PIOCPSINFO,&prusage) == -1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_FILE_READ,"Unable to access system file %s to get memory usage data",file);
  *mem = (PetscLogDouble)prusage.pr_byrssize;
  close(fd);

#elif defined(PETSC_USE_SBREAK_FOR_SIZE)

  *mem = (PetscLogDouble)(8*fd - 4294967296); /* 2^32 - upper bits */

#elif defined(PETSC_USE_PROC_FOR_SIZE) && defined(PETSC_HAVE_GETPAGESIZE)
  sprintf(proc,"/proc/%d/statm",(int)getpid());
  if (!(file = fopen(proc,"r"))) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_FILE_OPEN,"Unable to access system file %s to get memory usage data",proc);
  if (fscanf(file,"%d %d",&mm,&rss) != 2) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SYS,"Failed to read two integers (mm and rss) from %s",proc);
  *mem = ((PetscLogDouble)rss) * ((PetscLogDouble)getpagesize());
  err  = fclose(file);
  if (err) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SYS,"fclose() failed on file");

#elif defined(PETSC_HAVE_GETRUSAGE)
  getrusage(RUSAGE_SELF,&temp);
#if defined(PETSC_USE_KBYTES_FOR_SIZE)
  *mem = 1024.0 * ((PetscLogDouble)temp.ru_maxrss);
#elif defined(PETSC_USE_PAGES_FOR_SIZE) && defined(PETSC_HAVE_GETPAGESIZE)
  *mem = ((PetscLogDouble)getpagesize())*((PetscLogDouble)temp.ru_maxrss);
#else
  *mem = temp.ru_maxrss;
#endif


On May 13, 2014, at 12:13 PM, De Groof, Vincent Frans Maria <Vincent.De-Groof at uibk.ac.at> wrote:

> Hi,
> 
> 
> I'm investigating the performance of a few different direct solvers and I'd like to compare the memory requirements of the different solvers and orderings. I am especially interested in the memory usage necessary to store the factored matrix. 
> 
> I experimented with the PetscMemoryGetCurrentUsage and PetscMemoryGetMaximumUsage before and after KSPSolve. But these seem to return the memory usage on 1 process and not the total memory usage. Is this correct? I also noticed that the difference in maximum memory usage is very small before and after KSPSolve. Does it register the memory usage in external packages?
> 
> 
> 
> thanks,
> Vincent