[ZODB-Dev] [X-Post] Figure out bottle neck in a repoze.bfg based web app

Tue Jan 24 15:48:56 UTC 2012

On 24 January 2012 13:50, steve <steve at lonetwin.net> wrote:
> Hi All,
>
> I apologize for the cross-post but by this mail I simply hope to get a few
> pointers on how to narrow down to the problem I am seeing. I shall post to the
> relevant list if I have further questions.
>
> So here is the issue:
>
> Short description:
> I've got a repoze.bfg application running on top of zeo/zodb across multiple
> servers, served using mod_wsgi and it's showing bad resource usage (both high
> memory consumption as well as CPU usage). Are there any steps i can do to
> localise whether this is an issue with zeo/zodb/mod_wsgi configuration, and/or
> usage ?
>
> Long description:
>
> * I have a repoze.bfg (version 1.3) based app, which uses zodb (over zeo,
> version 3.10.2) as the backend and is served up using apache+mod_wsgi. All
> running on a minimal debian 6.0 based amazon instances.
>
> * The architecture is 1 zodb server and 4 app instances running on individual
> EC2 instances (all in the same availability zone). All of the instances are
> behind an amazon Elastic Load Balancer
>
> * At the web-server, we don't customize apache much (ie: we pretty much use the
> stock debian apache config). We use mod_wsgi (version 3.3-2) to serve the
> application in daemon mode, with the following parameters:
>
> WSGIDaemonProcess webapp user=appname threads=7 processes=4
> maximum-requests=10000 python-path=/path/to/virtualenv/eggs
>
> * The web app is the only thing that is served from these instances and we serve
> the static content for the using apache rather than the web app.
>
> * The zodb config on the db server looks like:
> <zeo>
>  address 8886
>  read-only false
>  invalidation-queue-size 1000
>  pid-filename $INSTANCE/var/ZEO.pid
>  # monitor-address 8887
>  # transaction-timeout SECONDS
> </zeo>
>
> <blobstorage 1>
>  <filestorage>
>    path $INSTANCE/var/webapp.db
>  </filestorage>
>  blob-dir $INSTANCE/var/blobs
> </blobstorage>
>
> * The zeo connection string (for repoze.zodbconn-0.11) is:
>
> zodb_uri = zeo://<zodb server
> ip>:8886/?blob_dir=/path/to/var/blobs&shared_blob_dir=false&connection_pool_size=50&cache_size=1024MB&drop_cache_rather_verify=true
>
> (Note: the drop_cache_rather_verify=true is for faster startups)
>
> Now with this, on live we have typical load such as:
> top - 13:34:54 up 1 day,  8:22,  2 users,  load average: 11.87, 8.75, 6.37
> Tasks:  85 total,   2 running,  82 sleeping,   0 stopped,   1 zombie
> Cpu(s): 81.1%us,  6.7%sy,  0.0%ni, 11.8%id,  0.0%wa,  0.0%hi,  0.1%si,  0.2%st
> Mem:  15736220k total,  7867340k used,  7868880k free,   283332k buffers
> Swap:        0k total,        0k used,        0k free,  1840876k cached
>
>  PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
>  5079 appname   21   0 1587m 1.2g 6264 S   77  8.1   9:23.86 apache2
>  5065 appname   20   0 1545m 1.2g 6272 S   95  7.9   9:31.24 apache2
>  5144 appname   20   0 1480m 1.1g 6260 S   86  7.4   5:49.92 apache2
>  5127 appname   20   0 1443m 1.1g 6264 S   94  7.2   7:13.10 apache2
> ....
> ....
> ....
>
> As you can see that very high load avg. and the apache processes spawned for
> mod_wsgi (identifiable because of the user whose context they run under) consume
> about 1.2Gs resident memory each.
>
> With a constant load like this, the app. response progressively degrades. We've
> tried to tweak the number of processes, the cache_size in the zeo connection
> string but all to no avail. So, now rather than shoot in the dark, I would
> appreciate suggestions on how I might be able to isolate the bottle-neck in the
> stack.
>
> One thing to note is that is high load and memory usage is only seen on the
> production instances. When we test the app. using ab or funkload on a similar
> setup (2 app instances instead of 4), we do not see this problem.
>
> Any pointers/comments would be appreciated.

(Following up only on zodb-dev as I'm not subscribed to the other lists.)

I'm guessing, but I suspect your load tests may only be reading from
the ZODB so you rarely see any cache misses.

The most important tuning paramaters for ZODB in respect to memory
usage are the number of threads and the connection_cache_size. The
connection_cache_size controls the number of persistent objects kept
live in the interpreter at a time. It's a per-connection setting and
as each thread needs its own connection. Memory usage increases
proportionally to connection_cache_size * number of threads. Most
people use either one or two threads per process with the ZODB. I know
plone.recipe.zope2instance defaults to two threads per process, though
I think this is only to avoid locking up in the case of Plone being
configured to load an RSS feed from itself.

The Python Global Interpreter Lock prevents threads from running
concurrently, so with ZEO running so many threads per process is
likely to be counter-productive. Try with one or two threads and
perhaps up the connection_cache_size (though loading from the zeo
cache is very quick you must ensure your working set fits in the
connection cache or else you'll be loading the same objects every
request). If your CPU usage goes down a lot it means your spending
time waiting for objects to be loaded in the connection and you might
want to run a couple more processes than you have cpus if you are
running one thread per process.

I also suggest you check that your mod_wsgi config is correctly
specifying WSGIProcessGroup, see:
http://www.martinaspeli.net/articles/update-repoze-under-mod-wsgi-is-not-slow
and http://code.google.com/p/modwsgi/wiki/ConfigurationDirectives#WSGIDaemonProcess.

To get more information about what is going on, try porting
https://github.com/collective/collective.stats to work as WSGI
middleware instead of hooking into Zope2's ZPublisher.

Laurence