Re: Unused Unicorn processes

Re: Unused Unicorn processes

[Konstantin Gredeskoul]

Greetings!

I have a question on optimal # of unicorn worker processes.

We are running Unicorn 4.3.1 + Rails 3.2.6 (without threading), on
ruby 1.9.3-p194, hosted on SmartOS/Joyent.

At the moment, unicorns are configured to start 30 worker processes. I
know this is a lot, and I am going to reduce this number. But in
trying to figure out what is a more appropriate number of workers to
run, I noticed something interesting that I couldn't explain.

If I look at the process table on each machine (see top output below),
I notice that some unicorn processes are heavily used (and have
accumulated longer CPU times, as well as have grown their RAM usage
since boot time), but other processes (at the bottom of the top
output) appear to potentially not having been used at all.  There are
several processes with RSS size of 143Mb, which I believe is unicorn
size before it processes any requests.

What I am gathering from this, is that only 16 unicorn processes are
actually processing requests, while the rest are just sitting there
idle.

Is this expected behavior?

This Joyent SmartMachine can burst up to 8 cores. Given that our web
requests spend only 80% of their time in ruby, I figured we could run
10 unicorn processes for maximum efficiency.  However seeing that 16
are actually used I am curious whether 16 is actually a better number.

Any suggestions appreciated!

load averages:  2.07,  1.97,  2.09;                    up 14+01:56:
58 processes: 53 sleeping, 5 on cpu
CPU states: 77.1% idle, 18.2% user,  4.7% kernel,  0.0% iowait,  0.
Memory: 8192M phys mem, 3157M free mem, 16G total swap, 11G free sw

   PID USERNAME LWP PRI NICE  SIZE   RES STATE    TIME    CPU COMMAND
 24800 wanelo    18  59    0  337M  313M sleep   17:21  2.05% ruby
 24788 wanelo    18  59    0  223M  200M cpu/14  16:34  3.18% ruby
 24787 wanelo    18  59    0  221M  197M sleep   14:55  1.89% ruby
 24776 wanelo    18  59    0  217M  193M cpu/2   18:17  3.45% ruby
 24804 wanelo    18  59    0  216M  192M cpu/8    7:54  0.33% ruby
 24785 wanelo    18  59    0  211M  187M sleep   14:33  0.75% ruby
 24783 wanelo    10  59    0  209M  185M sleep   16:09  2.16% ruby
 24775 wanelo    18  59    0  205M  181M sleep    5:24  0.03% ruby
 24777 wanelo    10  59    0  205M  181M cpu/11   3:53  1.54% ruby
 24799 wanelo    18  59    0  204M  179M sleep    0:42  0.00% ruby
 24803 wanelo    18  59    0  202M  177M sleep    2:28  0.35% ruby
 24802 wanelo     3  59    0  195M  168M sleep    1:45  0.01% ruby
 24778 wanelo     3  59    0  192M  165M sleep    1:06  0.01% ruby
 24801 wanelo     3  59    0  189M  162M sleep    0:50  0.01% ruby
 24798 wanelo     3  59    0  188M  161M sleep    0:30  0.00% ruby
 24797 wanelo     3  59    0  187M  159M sleep    0:19  0.00% ruby
--- unicorns below this line do not appear to be used
 24795 wanelo     3  59    0  173M  142M sleep    0:08  0.00% ruby
 24792 wanelo     3  59    0  172M  142M sleep    0:07  0.00% ruby
 24779 wanelo     3  59    0  172M  143M sleep    0:05  0.00% ruby
 24796 wanelo     3  59    0  172M  142M sleep    0:12  0.00% ruby
 24790 wanelo     3  59    0  172M  142M sleep    0:06  0.00% ruby
 24793 wanelo     3  59    0  172M  142M sleep    0:06  0.00% ruby
 24781 wanelo     3  59    0  172M  142M sleep    0:05  0.00% ruby
 24791 wanelo     3  59    0  172M  142M sleep    0:04  0.00% ruby
 24794 wanelo     3  58    0  172M  142M sleep    0:10  0.00% ruby
 24784 wanelo     3  59    0  172M  142M sleep    0:04  0.00% ruby
_______________________________________________
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

Re: Unused Unicorn processes

[Eric Wong]

Konstantin Gredeskoul <kig@wanelo.com> wrote:
> Greetings!
> 
> I have a question on optimal # of unicorn worker processes.
> 
> We are running Unicorn 4.3.1 + Rails 3.2.6 (without threading), on
> ruby 1.9.3-p194, hosted on SmartOS/Joyent.
> 
> At the moment, unicorns are configured to start 30 worker processes. I
> know this is a lot, and I am going to reduce this number. But in
> trying to figure out what is a more appropriate number of workers to
> run, I noticed something interesting that I couldn't explain.
> 
> If I look at the process table on each machine (see top output below),
> I notice that some unicorn processes are heavily used (and have
> accumulated longer CPU times, as well as have grown their RAM usage
> since boot time), but other processes (at the bottom of the top
> output) appear to potentially not having been used at all.  There are
> several processes with RSS size of 143Mb, which I believe is unicorn
> size before it processes any requests.
> 
> What I am gathering from this, is that only 16 unicorn processes are
> actually processing requests, while the rest are just sitting there
> idle.
> 
> Is this expected behavior?

It's not _unexpected_ if your load is low.  The OS scheduler does the
load balancing itself as unicorn uses non-blocking accept().  The
scheduler may choose busier processes because they're likely to be
hotter in cache/memory.

So you're probably getting better performance than if you had perfectly
balanced traffic.

Another possible explanation is something is killing some workers
(timeout, or crash buts, check stderr logs to confirm).  But even in a
situation where workers never die, it's perfectly normal to see workers
that never/rarely serve traffic.

> This Joyent SmartMachine can burst up to 8 cores. Given that our web
> requests spend only 80% of their time in ruby, I figured we could run
> 10 unicorn processes for maximum efficiency.  However seeing that 16
> are actually used I am curious whether 16 is actually a better number.

It depends what your app is waiting on.

The only rule is your machine should not be swap thrashing under
peak load.

Extra workers won't be detrimental as long as you:
1) have enough memory
2) have enough backend resources (e.g DB connections)

I know of deployments that run 30 processes/core because the app spends
much time waiting on slow HTTP/DB requests[1]

But if your app is completely bound by resources local to a machine (no
external DB/HTTP/memcached/redis/etc requests), then a 1:1 worker:core
(or even 1:1 worker:disk) relationship will work, too.



[1] Arguably a multi-threaded or non-blocking server such as Rainbows!
would be most efficient of machine resources if you're waiting on HTTP
calls, but the developers decided human time was more important and
did not want to worry about thread-safety.
_______________________________________________
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

Re: Unused Unicorn processes

[Konstantin Gredeskoul]

Thanks Eric!  Appreciate well thought out answers.

We are actually using Rainbows also, in an project where long
server-side HTTP calls are part of the application design and are
necessary, and it's working out really well.

Our main web app mostly spends it's time in CPU cycles serving web
pages (about 80% of the request time), and 20% in database operations.

Thanks again
K

On Tue, Aug 21, 2012 at 2:11 AM, Eric Wong <normalperson@yhbt.net> wrote:
> Konstantin Gredeskoul <kig@wanelo.com> wrote:
>> Greetings!
>>
>> I have a question on optimal # of unicorn worker processes.
>>
>> We are running Unicorn 4.3.1 + Rails 3.2.6 (without threading), on
>> ruby 1.9.3-p194, hosted on SmartOS/Joyent.
>>
>> At the moment, unicorns are configured to start 30 worker processes. I
>> know this is a lot, and I am going to reduce this number. But in
>> trying to figure out what is a more appropriate number of workers to
>> run, I noticed something interesting that I couldn't explain.
>>
>> If I look at the process table on each machine (see top output below),
>> I notice that some unicorn processes are heavily used (and have
>> accumulated longer CPU times, as well as have grown their RAM usage
>> since boot time), but other processes (at the bottom of the top
>> output) appear to potentially not having been used at all.  There are
>> several processes with RSS size of 143Mb, which I believe is unicorn
>> size before it processes any requests.
>>
>> What I am gathering from this, is that only 16 unicorn processes are
>> actually processing requests, while the rest are just sitting there
>> idle.
>>
>> Is this expected behavior?
>
> It's not _unexpected_ if your load is low.  The OS scheduler does the
> load balancing itself as unicorn uses non-blocking accept().  The
> scheduler may choose busier processes because they're likely to be
> hotter in cache/memory.
>
> So you're probably getting better performance than if you had perfectly
> balanced traffic.
>
> Another possible explanation is something is killing some workers
> (timeout, or crash buts, check stderr logs to confirm).  But even in a
> situation where workers never die, it's perfectly normal to see workers
> that never/rarely serve traffic.
>
>> This Joyent SmartMachine can burst up to 8 cores. Given that our web
>> requests spend only 80% of their time in ruby, I figured we could run
>> 10 unicorn processes for maximum efficiency.  However seeing that 16
>> are actually used I am curious whether 16 is actually a better number.
>
> It depends what your app is waiting on.
>
> The only rule is your machine should not be swap thrashing under
> peak load.
>
> Extra workers won't be detrimental as long as you:
> 1) have enough memory
> 2) have enough backend resources (e.g DB connections)
>
> I know of deployments that run 30 processes/core because the app spends
> much time waiting on slow HTTP/DB requests[1]
>
> But if your app is completely bound by resources local to a machine (no
> external DB/HTTP/memcached/redis/etc requests), then a 1:1 worker:core
> (or even 1:1 worker:disk) relationship will work, too.
>
>
>
> [1] Arguably a multi-threaded or non-blocking server such as Rainbows!
> would be most efficient of machine resources if you're waiting on HTTP
> calls, but the developers decided human time was more important and
> did not want to worry about thread-safety.
> _______________________________________________
> 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



-- 

Konstantin Gredeskoul
CTO :: Wanelo Inc
cell: (415) 265 1054
_______________________________________________
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