Re: High CPU Utilization

On Mar 20, 2009, at 4:29 PM, Scott Marlowe wrote:

> On Fri, Mar 20, 2009 at 2:26 PM, Joe Uhl <joeuhl@gmail.com> wrote:
>> On Mar 17, 2009, at 12:19 AM, Greg Smith wrote:
>>
>>> On Tue, 17 Mar 2009, Gregory Stark wrote:
>>>
>>>> Hm, well the tests I ran for posix_fadvise were actually on a
>>>> Perc5 --
>>>> though
>>>> who knows if it was the same under the hood -- and I saw better
>>>> performance
>>>> than this. I saw about 4MB/s for a single drive and up to about
>>>> 35MB/s
>>>> for 15
>>>> drives. However this was using linux md raid-0, not hardware raid.
>>>
>>> Right, it's the hardware RAID on the Perc5 I think people mainly
>>> complain
>>> about.  If you use it in JBOD mode and let the higher performance
>>> CPU in
>>> your main system drive the RAID functions it's not so bad.
>>>
>>> --
>>> * Greg Smith gsmith@gregsmith.com http://www.gregsmith.com
>>> Baltimore, MD
>>
>> I have not yet had a chance to try software raid on the standby
>> server
>> (still planning to) but wanted to follow up to see if there was any
>> good way
>> to figure out what the postgresql processes are spending their CPU
>> time on.
>>
>> We are under peak load right now, and I have Zabbix plotting CPU
>> utilization
>> and CPU wait (from vmstat output) along with all sorts of other
>> vitals on
>> charts.  CPU utilization is a sustained 90% - 95% and CPU Wait is
>> hanging
>> below 10%.  Since being pointed at vmstat by this list I have been
>> watching
>> CPU Wait and it does get high at times (hence still wanting to try
>> Perc5 in
>> JBOD) but then there are sustained periods, right now included,
>> where our
>> CPUs are just getting crushed while wait and IO (only doing about
>> 1.5 MB/sec
>> right now) are very low.
>>
>> This high CPU utilization only occurs when under peak load and when
>> our JDBC
>> pools are fully loaded.  We are moving more things into our cache and
>> constantly tuning indexes/tables but just want to see if there is
>> some
>> underlying cause that is killing us.
>>
>> Any recommendations for figuring out what our database is spending
>> its CPU
>> time on?
>
> What does the cs entry on vmstat say at this time?  If you're cs is
> skyrocketing then you're getting a context switch storm, which is
> usually a sign that there are just too many things going on at once /
> you've got an old kernel things like that.

cs column (plus cpu columns) of vmtstat 1 30 reads as follows:

cs    us  sy id wa
11172 95  4  1  0
12498 94  5  1  0
14121 91  7  1  1
11310 90  7  1  1
12918 92  6  1  1
10613 93  6  1  1
9382  94  4  1  1
14023 89  8  2  1
10138 92  6  1  1
11932 94  4  1  1
15948 93  5  2  1
12919 92  5  3  1
10879 93  4  2  1
14014 94  5  1  1
9083  92  6  2  0
11178 94  4  2  0
10717 94  5  1  0
9279  97  2  1  0
12673 94  5  1  0
8058  82 17  1  1
8150  94  5  1  1
11334 93  6  0  0
13884 91  8  1  0
10159 92  7  0  0
9382  96  4  0  0
11450 95  4  1  0
11947 96  3  1  0
8616  95  4  1  0
10717 95  3  1  0

We are running on 2.6.28.7-2 kernel.  I am unfamiliar with vmstat
output but reading the man page (and that cs = "context switches per
second") makes my numbers seem very high.

Our sum JDBC pools currently top out at 400 connections (and we are
doing work on all 400 right now).  I may try dropping those pools down
even smaller. Are there any general rules of thumb for figuring out
how many connections you should service at maximum?  I know of the
memory constraints, but thinking more along the lines of connections
per CPU core.