pyode-user

[Pyode-user] memory leaks

[Chris B. <chr...@gm...>]

Hi, the following code appears to leak Body and dict objects. Memprof
is available from 'svn co
http://svn.navi.cx/misc/trunk/python/memprof' (if anyone has a better
way of profiling python memory usage let me know!).


import ode

import memprof
sampler =3D memprof.Sampler()

sampler.run()
for i in range(1000):
    w =3D ode.World()
    body =3D ode.Body(w)
    # creating geom leads to 'dict' object leaks (_geom_c2py_lut ?)
    geom =3D ode.GeomCCylinder()
    # setBody leads to 'Body' object leaks
    geom.setBody(body)
    # no 'Body' leak if we set it back to None
    #geom.setBody(None)
    sampler.run()


I've looked at the pyode code. It appears that nothing is ever removed
from the global dictionary _geom_c2py_lut. Should this be done in geom
__dealloc__ ?

Similarly should geomobject __dealloc__ do 'del self.body'? Maybe it's
not necessary and geoms are being removed, it would appear that they
aren't from the profile, and yet what else other than the geom.body
reference could be keeping Body objects alive?

Thanks,
Chris

[Pyode-user] Re: memory leaks

[Chris B. <chr...@gm...>]

On 01/08/05, Chris Bainbridge <chr...@gm...> wrote:
> I've looked at the pyode code. It appears that nothing is ever removed
> from the global dictionary _geom_c2py_lut. Should this be done in geom
> __dealloc__ ?

The reference from _geom_c2py_lut to the Geom objects keeps them
alive, even with garbage collection. I solved this by doing
'ode._geom_c2py_lut.clear()' at the end of each simulation. There
should probably be a re-init function in the ode module itself.

Re: [Pyode-user] memory leaks

[Matthias B. <ba...@ir...>]

Chris Bainbridge wrote:
> Hi, the following code appears to leak Body and dict objects. Memprof
> [...]
> I've looked at the pyode code. It appears that nothing is ever removed
> from the global dictionary _geom_c2py_lut. Should this be done in geom
> __dealloc__ ?

The _geom_c2py_lut is a dictionary that serves as a lookup table to 
convert the geom pointer received from ODE into the appropriate Python 
geom object (which has to be done during collisions).
So when a geom is created it adds itself to the dictionary and indeed, 
it is never removed from there.
The geom cannot remove itself in __dealloc__() because as long as 
there's a reference in the dictionary it is not destroyed and 
__dealloc__() is never called. The solution would be weak references but 
at the time I wrote the stuff Pyrex didn't support this. I contacted the 
author, but I believe the current version of Pyrex still doesn't support 
it (I haven't checked explicitly but at least I never noticed anything 
in the changelogs).

As you said in your other mail, the workaround is to clear the 
dictionary manually when you're creating a new simulation scene.

> Similarly should geomobject __dealloc__ do 'del self.body'? 

Doesn't Python do that itself? Hm, but on the other hand, it's declared 
using "cdef", so you'r probably right and we should manually delete it 
in __dealloc__...
Then it would be the same for self.space in the Geoms and for self.world 
in the Body and the Joints.

- Matthias -

Re: [Pyode-user] memory leaks

[Chris B. <chr...@gm...>]

On 04/08/05, Matthias Baas <ba...@ir...> wrote:
> > Similarly should geomobject __dealloc__ do 'del self.body'?
>=20
> Doesn't Python do that itself? Hm, but on the other hand, it's declared
> using "cdef", so you'r probably right and we should manually delete it
> in __dealloc__...
> Then it would be the same for self.space in the Geoms and for self.world
> in the Body and the Joints.

No need - I'm clearing _geom_c2py_lut and I no longer have Body
objects leaking, it was just the lut->geom->body reference that kept
them alive.