[IronPython] IronPython with CherryPy through WSGI Memory issue

[Can Gencer]

On Thu, Apr 8, 2010 at 7:32 PM, Jeff Hardy <jdhardy at gmail.com> wrote:
> On Thu, Apr 8, 2010 at 10:38 AM, Can Gencer <cgencer at gmail.com> wrote:
>> I've noticed there was a leak in the NWSGI implementation regarding
>> the input/output streams, i.e. they were never GC'd. It could be due
>> to the way the stream -> file conversion is done.
>
> Wouldn't surprise me; that's kind of an ugly piece of code. There's a
> few spots where NWSGI doesn't interact with CodeContexts correctly,
> because there wasn't really a good story around them - now I'd just
> use SharedContext, assuming I can *find* it again :).
>
> But there must be something else, in NWSGI or IronPython, to account
> for 35k/request.
>

I did my debugging like this..

I set up a url such as http://server/gc where if you would go there it
would call GC.Collect() and GC.WaitForPendingFinalizers() a few times

Then did a few requests, and call the GC url and use windbg with
!dumpheap -stat and write down the number and size of the largest
objects. Do a few more requests, call the GC url again and use windbg
again with the stat, and then use !dumpheap -mt to get all instances
of the type that is growing in size most, and then randomly call
!gcroot on some to find where the reference is being held.

/Can