[petsc-users] proper Fortran usage for PetscMemoryGetMaximumUsage
Barry Smith
bsmith at mcs.anl.gov
Fri Jul 3 14:57:13 CDT 2015
>>> call PetscMemorySetGetMaximumUsage(ierr)
>>>
>>>
>>> .... some code with allocate, MatCreate, MatsetValues,...
>>>
>>> call PetscMemoryGetMaximumUsage(mem,ierr)
I should have realized this sooner. Your code is behaving as expected. The actual recording of maximum memory usage takes place when each object is destroyed (this gives the high water mark before the object is destroyed). Since you didn't destroy any objects since you turned on PetscMemorySetGetMaximumUsage it still has zero as the maximum usage. You need to call PetscMemoryGetMaximumUsage immediately after you destroy the object you care about.
Barry
> On Jul 3, 2015, at 1:35 PM, Anthony Haas <aph at email.arizona.edu> wrote:
>
> Hi Barry,
>
> I tried what you suggested. When I use the flag -memory_info, PetscMemoryGetMaximumUsage returns a nonzero value. If I do not use -memory_info, PetscMemoryGetMaximumUsage returns a zero value on all processes. Regarding PetscMemoryGetCurrentUsage, it returns a nonzero value with or without the -memory_info flag. As for the debugger, I set a breakpoint at the function location (break PetscMemoryGetMaximumUsage), then use run. When the program reaches the breakpoint, I use step and then print to examine mem (see below).
>
> Thanks,
>
> Anthony
>
>
> PetscMemoryGetMaximumUsage (mem=0x7fffffffcca0)
> at /home/anthony/LIB/petsc-3.6.0/src/sys/memory/mem.c:153
> 153 if (!PetscMemoryCollectMaximumUsage) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"To use this function you must first call PetscMemorySetGetMaximumUsage()");
> (gdb)
> 154 *mem = PetscMemoryMaximumUsage;
> (gdb) print *mem
> $1 = 6.9533558072018098e-310
> (gdb) step
> 155 PetscFunctionReturn(0);
> (gdb)
> 156 }
> (gdb)
> petscmemorygetmaximumusage_ (mem=0x7fffffffcca0, __ierr=0x7fffffffcd74)
> at /home/anthony/LIB/petsc-3.6.0/src/sys/memory/ftn-auto/memf.c:52
> 52 }
> (gdb)
> petsc::setupsolvegevp (nx=25, ny=25, nxy=625, alpha=0,
> beta=3.1415926535897931, re=1000) at module_petsc.F90:367
> 367 call MPI_Barrier(MPI_COMM_WORLD,ierr);CHKERRQ(ierr)
> (gdb) print mem
> $2 = 0
>
>
>
>
>
>
>
>
>
>
>
> On 07/02/2015 06:10 PM, Barry Smith wrote:
>>> On Jul 2, 2015, at 7:03 PM, Anthony Paul Haas <aph at email.arizona.edu> wrote:
>>>
>>> Hi Barry,
>>>
>>> Here it is. Note that the options -memory_info and -malloc_log provide proper data but I wanted to monitor some operations more closely.
>> That's odd, they all use the same mechanism.
>>
>> Run on one processor in the debugger use the PETSc command line option -start_in_debugger noxterm
>>
>> then put a breakpoint in PetscMemoryGetMaximumUsage and see what the values are when it gets there.
>>
>> Also run the same code with the option -memory_info maybe the PetscMemorySetGetMaximumUsage() isn't working, though it should,
>>
>> Barry
>>
>>> Thanks,
>>>
>>> Anthony
>>>
>>> On Thu, Jul 2, 2015 at 4:49 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>>>
>>> It is PetscLogDouble which is an alias for real*8.
>>>
>>> Only certain systems give easy access to the memory usage. Send the file $PETSC_ARCH/include/petscconf.h and I'll check if yours is suppose to be able to.
>>>
>>> Barry
>>>
>>>
>>>> On Jul 2, 2015, at 4:51 PM, Anthony Paul Haas <aph at email.arizona.edu> wrote:
>>>>
>>>> Hi,
>>>>
>>>> Could you please tell me what is the proper Fortran usage for PetscMemoryGetMaximumUsage? Do I need to define mem (see below) as PetscInt or PetscLogDouble? I tried both and in both case I got zero memory for all the processes.
>>>>
>>>> I did:
>>>>
>>>> 1st try: PetscLogDouble mem
>>>> 2nd try: PetscInt mem
>>>>
>>>> call SlepcInitialize(PETSC_NULL_CHARACTER,ierr)
>>>>
>>>> call MPI_COMM_SIZE(MPI_COMM_WORLD,size,ierr)
>>>> call MPI_COMM_RANK(MPI_COMM_WORLD,rank,ierr)
>>>>
>>>> call PetscMemorySetGetMaximumUsage(ierr)
>>>>
>>>>
>>>> .... some code with allocate, MatCreate, MatsetValues,...
>>>>
>>>> call PetscMemoryGetMaximumUsage(mem,ierr)
>>>>
>>>> write(*,*)'On Processor',rank,'memory:',mem
>>>>
>>>>
>>>> Thanks,
>>>>
>>>> Anthony
>>>
>>> <petscconf.h>
>
More information about the petsc-users
mailing list