Francois Tigeot <ftig...@wolfpond.org> wrote: > Hi Mindaugas, > > On Mon, Aug 27, 2012 at 04:11:59PM +0100, Mindaugas Rasiukevicius wrote: > > > > I am CC-ing tech-kern as well. > > Good. I was not sure in which of the numerous NetBSD lists I should have > posted to ;) > > > > On Mon, Aug 27, 2012 at 11:31:30AM +0100, Mindaugas Rasiukevicius > > > wrote: > > > > Francois Tigeot <ftig...@wolfpond.org> wrote: > > > > > > You're right, it's much better with a Unix socket. NetBSD follows the > > > FreeBSD curve up to 24 clients (including the peak). > > > Sadly, it all goes downhill from there. > > > > Good to know. Do you have a graph somewhere? > > I took some time to run an almost complete serie with NetBSD+AF_UNIX and > update the pdf document (see page 3). > Get it there: http://dl.wolfpond.org/benchs/Pg-benchmarks.2012-08.pdf
Thanks. Interestingly it is not exactly linear, it jumps at the peak. > > > Performance with 32 clients is on par with the initial 127.0.0.1 > > > results. Tests with 48 or more clients are still running and will > > > probably yield bad results too. > > > The system is almost completely unresponsive. Just seeing 'ls' appear > > > on the remote shell I use takes about 20 seconds with a long pause > > > between the two letters. > > > > This is somewhat unexpected. Do you use some unusual configuration or > > just a plain GENERIC kernel? Could you perhaps run pgbench under > > lockstat(8)? > > I'm using an almost-default installation with GENERIC. > If the kernel is before 15th of August, netbsd-6 still had DIAGNOSTIC option enabled by default, which would affect the performance. Did you use a kernel with the option disabled? Having lockstat output during the benchmark with 24 clients (peak) and separately with ~32 (when the system gets barely responsive) could help us to understand the problem better. -- Mindaugas
