[IronPython] HostCodeHeap leakage?

Mon Oct 18 11:09:49 CEST 2010

Ok, I finally succeeded in creating a simple reproduction for this problem.

The following code generates a 1000 methods (according to the ".NET CLR JIT"
performance counter), on Ipy 2.6.1 .Net 2.0

def test_method():

    for i in xrange(1000):

        def func(*a,**kw): pass

        func(some_parm = None)

test_method()

This does not happen if you call f without keyword params (using the *a
params is OK).

If this is indeed a bug, we would like to know how to fix it in the code
locally, if that is possible.

Also, I am interested in what Ipy flow creates this methods, since I wasn’t
able to find the function in the code that does this generation

Thanks

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Dino Viehland
*Sent:* Thursday, October 14, 2010 8:31 PM
*To:* Discussion of IronPython
*Subject:* Re: [IronPython] HostCodeHeap leakage?

So from an IronPython/DLR perspective the process should stabilize over
time.  That could take a while – as various Python functions get used
repeatedly we’ll switch from interpreting them to compiling them.  We’ll
also potentially produce new call site rules which are compiled.  That could
account for the increase of 16k->18k dynamic methods.  That’s a 12% increase
and could be a reasonable amount of it remained steady from then on.
Likewise the # of function codes seems stable but the _codeCount is rising.
That might mean that you’re defining new functions (via
exec/execfile/compile) and they’re getting collected - but we still have
their (dead) weak references in the code list.  Eventually that list should
get cleaned out when we hit context._nextCodeCleanup (which should be
greater than context._codeCount).

I would expect if you were to walk the entire PythonContext._allCodes linked
list that you’d see about half of the lists having a dead weak reference.
If your windbg-foo is up for writing the script to do this that’d be great
but mine is rusty enough I’d need to look it up in the documentation.

As for jitted code – all dynamic methods are collectible and any normal
RefEmit code is not collectible (there’s an option to make assemblies
collectible in .NET 4.0, but we don’t use it as we generally don’t generate
that many types).  Oh, we do also generate new types for subclasses but you
should see that NewTypeMaker._newTypes has a stable count over time because
we share these between types w/ common bases.

Closures, callbacks, generators should all be fine.

If you do .symfix, then .reload, in Windbg does “dt HostCodeHeap” show you
the fields of the HostCodeHeap structure?  Does !eeheap –loader give you the
address of the HostCodeHeap?  It might be useful to know what
m_allocationCount is on the heap overtime as if that’s relatively stable the
heap could be getting fragmented.  Looking at the CLR code I’m becoming more
and more convinced HostCodeHeap is only used for dynamic methods so if
that’s growing then I’m thinking we would appear to be leaking dynamic
methods.

The only way I can think of figuring out where the allocations are coming
from is to put a breakpoint on mscorwks!HostCodeHeap::AllocMemory_NoThrow
(hopefully this will show up in the public symbols).  That may be difficult
if you can’t attach the debugger to the server but if the issue is blocking
the server you could have a breakpoint here which does a stack trace and
continues execution so you could inspect the stack trace later.  I’d include
both “kb” and “!ClrStack” from sos in the stack trace.

2.7 shouldn’t really change this – whether .NET 4.0 would or not would
depend on if the CLR changed anything here.  But I’m not sure – I would
assume it wouldn’t.

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Idan Zaltzberg
*Sent:* Thursday, October 14, 2010 6:44 AM
*To:* Discussion of IronPython
*Subject:* Re: [IronPython] HostCodeHeap leakage?

Hi,

I've been looking on this for some time, and I'm still don’t understand some
things. Maybe I should begin by explaining our usage of Ipy (2.6.1 .NET 2.0)
a bit better.

We have a long running (stateful) application that we cannot simulate or run
with a debugger open (so no breakpoints).

However, since the application is run on a VM, we can take snapshots of it
and then open WinDbg instance and break in the middle of the application.

We do this a few hours after the application restarted and again after two
days so we can see the difference.

This way we saw that Jit Code Heap is increasing by a few hundred MB per
day, and the number of HostCodeHeap objects is increasing.

We also compared the performance counters for the two snapshots, and saw the
Jitted Code Bytes increased from 100MB to 862MB and the number of methods
jitted increased from 700K to 6.3M.

In the WinDbg we saw that the Jitted Code Heap size increases from 126MB to
424MB.

On the other hand, the object types you mentioned stay relatively the same:

·         DynamicMethods count went from 16K to 18K

·         FunctionCode count went from 4013 to 4025

·         The  _*code*Count field in the PythonContext went from 4447 to
7800

Here is what we don't understand:

1.       Is it normal for the application to keep jitting code and methods
forever? Should is stabilize?

2.       From the numbers I guess that some of the jitted code IS collected.
Which types are collectable and which are not? How can I tell which ones I
am using?

3.       Are there any specific patterns I should avoid to decrease
uncollectable code (or jitting in general). I am using a lot of closures,
callbacks and generators.

4.       What datatypes that are visible in WinDbg can I use to understand
if and why IronPython is generating uncollectable code?

5.       Is there a way to trace back the HostCodeHeap objects to my code
(or IronPython specific features)?

6.       Can I expect an improvement in these issues by moving to Ipy 2.7
and/or .Net 4.0?

Thanks

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Dino Viehland
*Sent:* Wednesday, October 13, 2010 10:01 PM
*To:* Discussion of IronPython
*Subject:* Re: [IronPython] HostCodeHeap leakage?

If you build from source you could set some breakpoints in AssemblyGen.cs in
the DefineType method.  You can also set one in DelegateUtils.cs
*and*DelegateHelpers.cs in DefineDelegateType.  I think those are all
the places
where we are creating uncollectible types.  If we’re continuously hitting
those breakpoints after you believe your app has reached steady state then
something is going wrong.

This code
http://www.koders.com/cpp/fid5CC8EACFCC85496B49B8CF83BD05AB36DE691E90.aspxleads
me to believe the HostCodeHeap might also be used for DynamicMethods.
If that is the case then the other place to look would be if FunctionCode
objects are being re-created repeatedly.  That will happen if there’s
exec/eval/compile calls which are happening and if those objects are being
kept alive then we could be growing the heap over time.

There’s also some complicated code which deals with keeping a list of all
code that is alive.  We do cleanup this list, and the list is a list of weak
references so it shouldn’t actually keep the code alive, but you could put
some breakpoints at FunctionCode.RegisterFunctionCode and
FunctionCode.CodeCleanup to see if that list is growing boundlessly (which
it would be if something was keeping code objects alive after an
exec/eval/compile).

Another place where code generation could be occurring would be w/ regexes.
If you are dynamically generating reg-exes, or executing a huge different
variety of them over time, and they’re compiled, then the compiled regexes
could be staying in memory.  There is a regex cache and you can clear it by
calling re.purge().  But it should only cache up to 100 regexes.

A final possible thing to investigate might be what happens if you throw
away the entire ScriptEngine instance.  Here you could try re-cycling the
ScriptEngine say every 6 hours and see if the problem goes away.  If that
fixes the problem then it’s likely that it is one of the things I mentioned
(or some other cache that’s per-runtime).  At least that would start to
narrow it down vs. some potentially global state (like the subtype list
which is shared across ScriptEngines).

That’s a bunch of different things to look at – hopefully it’ll give some
insight into what’s going on and help track down the issue.

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Idan Zaltzberg
*Sent:* Wednesday, October 13, 2010 12:12 AM
*To:* Discussion of IronPython
*Subject:* Re: [IronPython] HostCodeHeap leakage?

I tried what you suggested (changed setup.DebugMode = false;)

But still I get the same behavior:

The "Jit Code Heap" increases from about 17MB to 230MB in 2 days.

Is there a way to verify from the IronPython code that DebugMode is off?

Is there anything else I can do (other startup settings?) to
decrease/understand the increase in HostCodeHeap objects?

Thanks.

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Dino Viehland
*Sent:* Tuesday, October 05, 2010 6:59 PM
*To:* Discussion of IronPython
*Subject:* Re: [IronPython] HostCodeHeap leakage?

Yep, DebugMode is the same as –X:Debug.  In general I’d suggest making this
configurable somehow and only turn it on if you’re actually debugging.  It’s
unfortunate that we can’t offer both debugging & collectability but right
now that’s simply a limitation of the CLR and/or our lack of a separate VS
debug engine which can debug Python code.

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Idan Zaltzberg
*Sent:* Tuesday, October 05, 2010 9:09 AM
*To:* Discussion of IronPython
*Subject:* Re: [IronPython] HostCodeHeap leakage?

Im running using the engine from a hosting app.

We have these lines in the startup:

ScriptRuntimeSetup setup = new ScriptRuntimeSetup();
setup.DebugMode = true;

ScriptRuntime runtime = Python.CreateRuntime(setup.Options);

engine = runtime.GetEngine("py");

Is this is the same like –X:Debug?

You reckon this could be the cause?

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Dino Viehland
*Sent:* Tuesday, October 05, 2010 5:53 PM
*To:* Discussion of IronPython
*Subject:* Re: [IronPython] HostCodeHeap leakage?

My guess is that’s code in the JIT heap that’s building up but I’m not 100%
certain.  How is your code being executed?  Do you have the debug option (-D
or –X:Debug) enabled?  To support debug mode we need to produce
uncollectible code which could be building up.

*From:* users-bounces at lists.ironpython.com [mailto:
users-bounces at lists.ironpython.com] *On Behalf Of *Idan Zaltzberg
*Sent:* Tuesday, October 05, 2010 2:26 AM
*To:* Discussion of IronPython
*Subject:* [IronPython] HostCodeHeap leakage?

I am trying to find a memory/"performance" leak in an Ipy application.

Using WINDBG (!eeheap -loader), we noticed the that the LoaderHeap is
getting bigger (150MB increase per day). From the !eeheap output it seems
that the increase is due to HostCodeHeap (objects?).

As I understand these objects might be created by Ipy infra, is that right?

Is there anyway I can get more info on their content, or prevent them from
growing?

Thanks
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/ironpython-users/attachments/20101018/b884776b/attachment.html>