Newer kernels and CFS scheduler again

Hi All,

As I'll soon be looking at migrating some of our debian servers onto the new stable release, I've started doing a bit of basic pgbench testing.

Initially I've seen a little performance regression with higher concurrent clients when going from the 2.6.32 kernel to 3.2.14 (select only and tpc-b).  After trying the suggestions made by Shaun Thomas a while back (here: http://www.postgresql.org/message-id/50E4AAB1.9040902@optionshouse.com) and getting nowhere, I'm seeing big improvements instead increasing the

defaults for sched_min_granularity_ns and sched_wakeup_granularity_ns (As described here: https://www.kernel.org/doc/Documentation/scheduler/sched-design-CFS.txt) from debians defaults of 3000000 and 4000000 respectively.

On my initial test setup (which admittedly is far from cutting edge) of 2xE5320 / 32Gb the following seem pretty optimal:

kernel.sched_min_granularity_ns=9000000
kernel.sched_wakeup_granularity_ns=12000000

I've yet to do any testing on our larger machines, but as there have been a few posts here about performance vs newer kernels I was just wondering what other peoples findings are regarding CFS?

Glyn