On 12 April 2011 10:26, Barry Smith <[email protected]> wrote:

On Apr 11, 2011, at 10:02 PM, Lisandro Dalcin wrote:

...

After adding this test in petsc4py-dev/test/test_ksp.py :

   def testResetAndSolve(self):        self.ksp.reset()        self.testSolve()        self.ksp.reset()        self.testSolve()        self.ksp.reset()

it seems there is some memory bleeding:

[0]Total space allocated 256 bytes [ 0]80 bytes KSPSetUp_BCGSL() line 476 in /home/devel/petsc/dev/src/ksp/ksp/impls/bcgsl/bcgsl.c      [0]  KSPSetUp() line 184 in /home/devel/petsc/dev/src/ksp/ksp/interface/itfunc.c      [0]  KSPSolve() line 332 in /home/devel/petsc/dev/src/ksp/ksp/interface/itfunc.c

These are workspace arrays that are re-allocated at every KSPSetup_BCGSL call.

Not sure about the proper way to fix this, but I would deallocate the arrays at KSPReset_BCGSL. Can anyone help me with this?

   This could be a problem in many places: for example if eigenvalue computing is done with CG then I think those arrays will be reallocated.

Indeed. I'm not testing eigenvalue computing. I could try to add the options to my testsuite and see what happens.

  Previously I guess KSPSetUp_XXX() never had a reason to be called more than once, with reset now it is. So that means that all arrays allocated in KSPSetUp_XXX() should be freed in KSPReset_XXX()? If this is correct then it is easily fixed by moving them from KSPDestroy_XXX().

Well, In my latest push I was lazy enough as to not remove the PetscFree() call to Reset_XXX. After all, PetscFree() nullifies the pointer, so we are safe. If you want me to fix this, let me know. PS: Could any of you check BCGLS for a few "ksp->setupstage = KSP_SETUP_NEW;" lines? That really smells, such hack forces a full (not just in the subype) KSPSetup() in the next KSPSolve(). -- Lisandro Dalcin --------------- CIMEC (INTEC/CONICET-UNL) Predio CONICET-Santa Fe Colectora RN 168 Km 472, Paraje El Pozo 3000 Santa Fe, Argentina Tel: +54-342-4511594 (ext 1011) Tel/Fax: +54-342-4511169