Re: Query regarding postgres lock contention - Followup

On Wed, Oct 5, 2011 at 1:48 PM, Kevin Grittner
<Kevin(dot)Grittner(at)wicourts(dot)gov> wrote:
> Hamza Bin Sohail <hsohail(at)purdue(dot)edu> wrote:
>
>> My postgres version is 8.3.7
>
> Why such an old version?  Why exclude the available bug fixes?
>
> http://www.postgresql.org/support/versioning
>
>>> I am aware that lock contention can be checked with lockstat (and
>>> with pg_locks ? ) but I wanted to know if someone can tell me how
>>> much contention there would be for this database in a 16-core
>>> system vs a 4-core system. I just need a rough idea.
>
> How many database connections will be used?  If more than about
> twice the number of cores, you should probably be going through a
> transaction-based connection pool.
>
> With 16 cores, even with a properly configured connection pool, you
> will probably be on the edge of where spinlock contention starts
> eating significant CPU time.  With enough system RAM and proper
> tuning the hit should be fairly minor, I think.  It really gets bad
> at 32 cores, although that is being improved for next year's 9.2
> release.

I think that on write-heavy workloads (like pgbench) we bottleneck on
lightweight lock contention around 8 cores. :-(

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company