Re: How to find the culprit in server load spikes?

Moreno Andreo <moreno(dot)andreo(at)evolu-s(dot)it> wrote:

> I host a Postgresql server on Ubuntu 12.04 and I am facing server
> load spikes (if I run top, it goes up to 25-30 on a 4-core system)...
> In some cases, I have to restart potgresql service because users call us
> complaining of the slowness, but in some cases I can leave things on
> their way and I see that after a bunch of minutes (about 5-10) the
> situations drops to the normality (0.50-2 load).
>
> The problem is, as in the most cases, the I/O,

If you have confirmed that there is an I/O glut during these
episodes, it is probably that a cascade of dirty cache pages
caused the OS dirty pages to hit the vm.dirty_ratio percentage. If
you were seeing a high number for system CPU time the below would
probably not help.

Make sure you are using a storage system that has a persistent
cache configured for write-back (rather than write-through).
Reduce the OS vm.dirty_background_bytes setting to less than the
size of the persistent write cache. Make sure that vm.dirty_ratio
is at least 20, possibly higher. Configure the PostgreSQL
background writer to be more aggressive. If those don't do it,
reduce the size of shared_buffers.

--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company