Anyone interested in improving postgresql scaling?

[Robert Watson]

On Tue, 10 Apr 2007, Kevin Way wrote:

> Kris Kennaway wrote:
>> If so, then your task is the following:
>> 
>> Make SYSV semaphores less dumb about process wakeups.  Currently whenever 
>> the semaphore state changes, all processes sleeping on the semaphore are 
>> woken, even if we only have released enough resources for one waiting 
>> process to claim.  i.e. there is a thundering herd wakeup situation which 
>> destroys performance at high loads.  Fixing this will involve replacing the 
>> wakeup() calls with appropriate amounts of wakeup_one().
> Could this cause problem cause a situation where an 8-Core system was 50-75% 
> slower than an otherwise equivalent 2-Core system?
>
> I have a graph of my sysbench/pgsql results here:
>
>   http://blog.insidesystems.net/files/sysctl-pgsq-amd64-wtf.png
>
> As the graph shows, the 8-core system is about half the speed of the 2-core 
> system at 2 simultaneous threads, and it decays down to approximately 1/4 
> the speed of the 2-core system as the # of threads hits 5.
>
> All other (non-pgsql, non-sysv) tests came back approximately as expected, 
> but I'm left wondering if I did something wrong, or if 8 cpus are slower 
> than 2, when it comes to Postgres on currently available FreeBSD.
>
> Kevin Way Inside Systems, Inc.
>
> (Full detail of what I did is available at: 
> http://blog.insidesystems.net/articles/2007/04/09/what-did-i-do-wrong )

Would you be able to re-run these tests trying a recent 7.x kernel?  If so, 
make sure you build a non-debugging kernel, and try two variants: one with 
SCHED_4BSD, and one with SCHED_ULE.

Another experiment that would be interesting is to try using device.hints to 
disable various numbers of CPUs on your 8-core system, and see how the 
performance graph changes as the number of enabled CPUs changes.

Robert N M Watson
Computer Laboratory
University of Cambridge