python gc performance in large apps

[Robby Dermody]

Hey guys,

An update (along with a request for paid help at the end): Over the past 
week and a half I've improved the memory usage situation by quite a bit. 
Going to python 2.4, linux kernel 2.6, twisted 2.0 and altering some 
code have reduced the severity by an order of a magnitude. However, on a 
simulated call center environment (constant 50 conversations, switching 
every 300 seconds) the director component still consumes an average of 
1MB more per hour and the harvester is taking an average of 4MB more per 
hour. With the director, 2/3 of this is resident (1/3 in swap). With the 
harvester, about 88% of this is resident (~ 12% in swap).

After looking into things more, based on the helpful input I got from 
the very intelligent individuals who responded to the thread, I know a 
few more things:

-Neither component has any uncollectable trash problems. gc.garbage 
never has any items after a full collection (and there are never any 
uncollectable cycles).

-Running the harvester through valgrind under very light load showed 2 
minor memory leaks of my making in the "harvester core" (pyrex code). 
Both were fixed. Subsequent runs yielded nothing in my pyrex modules.

-len(gc.get_objects()) will show linearly increasing counts over time 
with the director (that match the curve of the director's memory usage 
exactly), but with the harvester, the object count doesn't increase over 
time (although memory usage does). This might mean that we are dealing 
with two separate problems on each component....uncollected object count 
growth on the director, and something else on the harvester. ...OR the 
harvester may have an object count growth problem as well, it might just 
be in a C module in a place not visible to gc.get_objects() ?

-I have some code that looks through gc.get_objects() to find the 10 
biggest lists and dicts, and I see no tell tale signs there (nothing 
growing out of control).

-Running the harvester on python 2.4 (and 2.5, same results) compiled 
with COUNT_ALLOCS reported some interesting results with lists and dicts 
that may explain some of the memory usage:
t = 0:
	list alloced: 19167, freed: 8231, max in use: 10937
	dict alloced: 47943, freed: 34108, max in use: 13837
t = 120 seconds (1st run after being fully initialized):
	list alloced: 2394620, freed: 17565, max in use: 2377056
	dict alloced: 2447968, freed: 67999, max in use: 2379969
t = ~20 hours:
	list alloced: 834032375, freed: 4625810, max in use: 829406570
	dict alloced: 845149422, freed: 15715727, max in use: 829433695

The numbers for the other types mostly looked normal, but these two kept 
growing after every subsequent call to sys.getcounts(). I have not yet 
run the director with COUNT_ALLOCS yet, due to a problem dynamically 
loading in the kinterbasdb (Firebird SQL interface) module.

-Memory usage is constant on both components when x conversations are 
started but never stopped. The code that handles that is just in the 
harvester, and is in pure C (in pyrex)... when a conversation is started 
or stopped, etc, python and pyrex code in both components runs, and 
things like twisted PB are used. This 1MB/hour and 4MB/hour increase is 
happening just by having the active convos stop and restart, but I can't 
see what would cause that kind of growth.

-Running on Python 2.5 (latest CVS HEAD) didn't change anything


At this point I think I've done all that I can do on my own with this, 
given my meager level of knowledge in this field :). I would be 
interested in retaining some professional help to see if, working 
together, we can trace down some likely culprits. I am also very 
interested in learning more on how to avoid these problems in the 
future, if they are my doing. I doubt we can remove all the memory 
problems, but if we can get this down by another 150 - 200%, that will 
be good enough for me.

With that said, I would be prepared to pay someone to help me out with 
this problem (and offer tips on improving the code... I am very 
interested in improving both my skills and the code's quality). The 
director and harvester programs can both run on a single box running 
Linux kernel 2.6, python 2.4. There aren't that many dependencies, and I 
have a simple loadtester program that can be used to stress test the 
environment. I've also already written a module for both components that 
logs a bunch of useful memory usage statistics out to disk at whatever 
interval desired.

I would be looking for an experienced python developer with the 
following knowledge:

-python internals, python "best practices"
-twisted framework
-pyrex (and python C API)
-previous experience fixing similar problems in python

If anyone is interested in helping out, or know someone who might be 
interested, please get back to me with your asking rate (hourly, or by 
the project if you prefer). We can go from there.

Also, as this project is starting to generate revenue, I will be hiring 
someone full time soon to take over primary development. This person 
doesn't need to be a python god. Someone with 2-3+ years of experience 
with python, subversion, etc. would be good enough, as long as they are 
a geek at heart, intelligent and adaptive, passionate about this kind of 
work and self motivated. Even a bright recent college grad with open 
source experience under his/her belt would be fine. If anyone knows of 
anyone who might love this kind of job, let me know.  This is truly a 
_very_ interesting product.

Robby


Robby Dermody wrote:
> Hey guys (thus begins a book of a post :),
> 
> I'm in the process of writing a commercial VoIP call monitoring and 
> recording application suite in python and pyrex. Basically, this 
> software sits in a VoIP callcenter-type environment (complete with agent 
> phones and VoIP servers), sniffs voice data off of the network, and 
> allows users to listen into calls. It can record calls as well. The 
> project is about a year and 3 months in the making and lately the 
> codebase has stabilized enough to where it can be used by some of our 
> clients. The entire project has about 37,000 lines of python and pyrex 
> code (along with 1-2K lines of unrelated java code).
> 
> Now, some disjointed rambling about the architecture of this software. 
> This software has two long-running server-type components. One 
> component, the "director" application, is written in pure python and 
> makes use of the twisted, nevow, and kinterbasdb libraries (which I 
> realize link to some C extensions). The other component, the 
> "harvester", is a mixture of python and pyrex, and makes use of the 
> twisted library, along with using the C libs libpcap and glib on the 
> pyrex end. Basically, the director is the "master" component. A single 
> director process interacts with users of the system through a web and/or 
> pygtk client application interface and can coordinate 1 to n harvesters 
> spread about the world. The harvester is the "heavy lifter" component 
> that sniffs the network traffic and sifts out the voice and signalling 
> data. It then updates the director of call status changes, and can 
> provide users of the system access to the data. It records the data to 
> disk as well. The scalibility of this thing is really cool: given a 
> single director sitting somewhere coordinating the list of agents, 
> multiple harvester can be placed anywhere there is voice traffic. A user 
> that logs into the director can end up seeing the activity of all of 
> these seperate voice networks presented like a single giant mesh.
> 
> Overall, I have been very pleased with python and the 3rd party 
> libraries that I use (twisted, nevow, kinterbasdb and pygtk). It is a 
> joy to program with, and I think the python community has done a fine 
> job. However, as I have been running the software lately and profiling 
> its memory usage, the one and only Big Problem I have seen is that of 
> the memory usage. Ideally, the server application(s) should be able to 
> run indefinitely, but from the results I'm seeing I will end up 
> exhausting the memory on a 2 GB machine in 2 to 3 days of heavy load.
> 
> Now normally I would not raise up an issue like this on this list, but 
> based on the conversations held on this list lately, and the work done 
> by Evan Jones (http://evanjones.ca/python-memory.html), I am led to 
> believe that this memory usage -- while partially due to some probably 
> leaks in my program -- is largely due to the current python gc. I have 
> some graphs I made to show the extent of this memory usage growth:
> 
> http://public.robbyd.fastmail.fm/iq-graph1.gif
> 
> http://public.robbyd.fastmail.fm/iq-graph-director-rss.gif
> 
> http://public.robbyd.fastmail.fm/iq-graph-harv-rss.gif
> 
> The preceding three diagrams are the result of running the 1 director 
> process and 1 harvester process on the same machine for about 48 hours. 
> This is the most basic configuration of this software. I was running 
> this application through /usr/bin/python (CPython) on a Debian 'testing' 
> box running Linux 2.4 with 2GB of memory and Python version 2.3.5. 
> During that time, I gathered the resident and virtual memory size of 
> each component at 120 second intervals. I then imported this data into 
> MINITAB and did some plots. The first one is a graph of the resident 
> (RSS) and virtual memory usage of the two applications. The second one 
> is a zoomed in graph of the director's resident memory usage (complete 
> with a best fit quadratic), and the 3rd one is a zoomed in graph of the 
> harvester's resident memory usage.
> 
> To give you an idea of the network load these apps were undergoing 
> during this sampling time, by the time 48 hours had passed, the 
> harvester had gathered and parsed about 900 million packets. During the 
> day there will be 50-70 agents talking. This number goes to 10-30 at night.
> 
> In the diagrams above, one can see the night-day separation clearly. At 
> night, the memory usage growth seemed to all but stop, but with the 
> increased call volume of the day, it started shooting off again. When I 
> first started gathering this data, I was hoping for a logarithmic curve, 
> but at least after 48 hours, it looks like the usage increase is almost 
> linear. (Although logarithmic may still be the case after it exceeds a 
> gig or two of used memory. :) I'm not sure if this is something that I 
> should expect from the current gc, and when it would stop.
> 
> Now, as I stated above, I am certain that at least some of this 
> increased memory usage is due to either un-collectable objects in the 
> python code, or memory leaks in the pyrex code (where I make some use of 
> malloc/free). I am working on finding and removing these issues, but 
> from what I've seen  with the help of gc UNCOLLECTABLE traces, there are 
> not many un-collectable reference issues at least. Yes, there are some 
> but definitely not enough to justify growth like I am seeing. The pyrex 
> side should not be leaking too much, I'm very good about freeing what I 
> allocate in pyrex/C land. I will be running that linked to a memory leak 
> finding library in the next few days. Past the code reviews I've done, 
> what makes me think that I don't have any *wild* leaks going on at least 
> with the pyrex code is that I am seeing the same type of growth patterns 
> in both apps, and I don't use any pyrex with the director. Yes, the 
> harvester is consuming much more memory, but it also does the majority 
> of the heavy lifting.
> 
> I am alright with the app not freeing all the memory it can between high 
> and low activity times, but what puzzles me is how the memory usage just 
> keeps on growing and growing. Will it ever stop?
> 
> What I would like to know if others on this list have had similar 
> problems with python's gc in long running, larger python applications. 
> Am I crazy or is this a real problem with python's gc itself? If it's a 
> python gc issue, then it's my opinion that we will need to enhance the 
> gc before python can really gain leverage as a language suitable for 
> "enterprise-class" applications. I have surprised many other programmers 
> that I'm writing an application like this in python/pyrex that works 
> just as well and even more efficiently than the C/C++/Java competitors. 
> The only thing I have left to show is that the app lasts as long between 
> restarts. ;)
> 
> 
> Robby