Re: Significant performance difference for postgres w/o sched_autogroup

[Andres Freund <andres@anarazel.de>]

On 2015-07-09 04:45:38 +0200, Mike Galbraith wrote:
> On Wed, 2015-07-08 at 17:45 +0200, Andres Freund wrote:
> > Workload:
> > 
> > postgresql (multi-process via shared memory SQL server) and benchmark
> > client (pgbench, multi threaded) running on the same server. Connected
> > using unix sockets.  The statements are relatively simple (~1.5ms on
> > average), forcing frequent back/forth between server processes and
> > pgbench threads.
> > 
> > I found that disabling sched_autogroup *significantly* reduces
> > throughput. Even when both server and client are started from the same
> > terminal and thus should be in the same group!
> > 
> > There's a significant difference in how %sys with autogroups
> > enabled/disabled. ~8% v ~27%. That sounds too much.
> 
> Seems reasonable to me.  1(tg)/2(tgs) > 1(task)/N(tasks), throughput is
> what the server can sustain on its given budget, larger budget means
> less client blockage, thus less %sys.

That'd make some sense if there were other stuff going on - but here the
same total budget in both cases leads to a 40% difference in throughput:

autogroup on: tps = 21329.219141; autogroup off: tps = 15006.317841

there's really nothing to do for the kernel or other tasks in the
background. It's a readonly workload, not requiring lots of memory, ...