Re: Fwd: Support for Soft Timeout in Unicorn

[Ryan Tomayko <r@tomayko.com>]

On Fri, Jun 18, 2010 at 2:48 PM, Eric Wong <normalperson@yhbt.net> wrote:
> Ryan Tomayko <r@tomayko.com> wrote:
>> On Fri, Jun 4, 2010 at 1:59 PM, Eric Wong <normalperson@yhbt.net> wrote:
>> > Chris Wanstrath <chris@ozmm.org> wrote:
>> >> That's what we do at GitHub. We're running Rails 2.2.2 and have a
>> >> custom config.ru, thanks to Unicorn:
>> >>
>> >> http://gist.github.com/424352
>> >
>> > By the way, how's the OobGC middleware working for you guys?
>>
>> We rolled out the OobGC middleware along with a basic RailsBench GC
>> config (RUBY_HEAP_MIN_SLOTS, etc.). Combined, they knocked about 20ms
>> (~15%) off the average response time across the site (real traffic).
>> The impact was significantly more for requests that allocate a lot of
>> objects -- as much as 50% decreases in response time for the worst
>> offenders. We saw no noticeable increase in CPU with OobGC set to run
>> every 10 requests, and a fair increase in CPU with OobGC set to run
>> every 5 requests.
>
> Cool.  Am I correct to assume the increased CPU usage at every 5
> requests wasn't worth any performance gains you might have had?

Yes. I don't have exact numbers on how much more CPU was being
utilized but it was definitely noticeable on my graphs and had very
little, if any, effect on overall response time beyond the savings at
the req/10 interval.

>> Because I rolled this stuff out somewhat non-scientifically, I've
>> always wondered how much OobGC contributed to the overall savings vs.
>> the RailsBench GC config. Disabling the OobGC middleware but leaving
>> the RailsBench GC config in place, I get the following graph:
>>
>> http://img.skitch.com/20100618-kihdc1cq6pjhq9rqftf8miuf6y.png
>>
>> So we're spending ~1ms request time in GC with OobGC, and ~10ms
>> request time in GC without it.
>
> Awesome.
>
>> Here's some system load graphs for the same time period just to show
>> that OobGC has no adverse effect when set to GC every 10 requests:
>>
>> http://img.skitch.com/20100618-qp7p8f6i2agbqbdnjbpigik1d9.png
>>
>> I assume the RailsBench GC patches improve the effect of OobGC
>> considerably by increasing the number of objects that can be allocated
>> between GC runs, allowing more of the GC work to be deferred to
>> in-between-requests time. Here's the RailsBench config we're using
>> today, for the record:
>>
>>     RUBY_HEAP_MIN_SLOTS=800000
>>     RUBY_HEAP_FREE_MIN=100000
>>     RUBY_HEAP_SLOTS_INCREMENT=300000
>>     RUBY_HEAP_SLOTS_GROWTH_FACTOR=1
>>     RUBY_GC_MALLOC_LIMIT=79000000
>>
>> This is only barely tuned for us. I stole most of the numbers from the
>> Twitter and 37signals examples.
>>
>> I've also experimented with tuning the GC specifically to take
>> advantage of OobGC:
>>
>> https://gist.github.com/87d574a19372c6043c5f
>>
>> # The following GC settings have been tuned for GitHub application web requests.
>> # Most settings are significantly higher than the example configs published by
>> # Twitter and 37signals. There's a couple reasons for this. First, the GitHub
>> # app has a memory footprint that's 3x-4x larger than the standard Rails app
>> # (roughly 200MB after first request compared to ~40MB-50MB). Second, because
>
> Yikes, 200MB after one request is a lot.  If you can easily find ways to
> cut that down, it should be more of a gain than the monster heap you've
> tried.

Indeed. We bring in *a lot* of libraries, some of which add
embarassingly large amounts of code and data at require time (aws-s3,
googlecharts - I'm looking at you).

>> # Unicorn is such an exceptional piece of software, we're able to schedule GC
>> # to run outside the context of requests so as not to effect response times.
>> # As such, we try to allocate enough memory to service 5 requests
>> without needing
>> # GC and then run GC manually immediately after each fifth request has been
>> # served but before the process starts accepting the next connection. The result
>> # is higher memory use (~300MB per Unicorn worker process on average) and a
>> # slight increase in CPU due to forced manual GC, but better response times.
>> # ...
>>
>> Unfortunately, the bigger heap seems to cause a largish increase in
>> the time needed for each GC, so the unicorn workers were spending too
>> much time between requests. CPU and RES increases were even more than
>> I'd expected. It also didn't eliminate in-request GC entirely as I'd
>> hoped.
>>
>> I eventually abandoned the idea -- even if I could get it to behave,
>> it's hardly worth the 1ms it would save. I mention it here because the
>> general approach might work in situations where the base heap size is
>> a bit smaller (say < 80MB) or perhaps I'm mistuning one of the
>> parameters.
>
> So in conclusion, OobGC itself works, but too large of a heap isn't
> worth it even for a memory hungry app.
>
> I suppose having too big of a heap means it can fragment more badly.
> Making GC run more often on a smaller heap can and give similar (or
> maybe better) performance.  At best you'll get diminishing returns as
> you seem to have concluded.

That sounds like an extremely plausible explanation.

> I have no doubt the Railsbench GC patches help.  Even with small apps,
> being able being able to set a linear growth factor on long-running
> servers is awesome.

Sure is.

Thanks,
Ryan

> Thanks for sharing this!
>
> --
> Eric Wong
> _______________________________________________
> Unicorn mailing list - mongrel-unicorn@rubyforge.org
> http://rubyforge.org/mailman/listinfo/mongrel-unicorn
> Do not quote signatures (like this one) or top post when replying
>
_______________________________________________
Unicorn mailing list - mongrel-unicorn@rubyforge.org
http://rubyforge.org/mailman/listinfo/mongrel-unicorn
Do not quote signatures (like this one) or top post when replying