On Fri, 10 Apr 2009, Glyn Astill wrote:
> So it was only for connections over a unix socket, but wow; it's still
> an ongoing issue.
The problem is actually with pgbench when running on a UNIX socket, not
with the PostgreSQL server itself. On my tests, the actual database
server itself seems to work just as well or better on later kernels that
use the new scheduler than the older scheduler did.
Basically, if all these apply:
1) You are running pgbench
2) You're running a quick statement, such as a simple select, that gives
>10000TPS or so
3) Connecting via UNIX socket
4) Clients > around 10
5) Linux kernel >=2.6.23 (which means CFS as the scheduler)
6) The CFS features are at their defaults (SCHED_FEAT_SYNC_WAKEUPS is on)
You'll get weird results. Change any of those and things are still fine.
--
* Greg Smith gsmith@gregsmith.com http://www.gregsmith.com Baltimore, MD