Messages in this thread | | | Subject | Re: load balancing regression since commit 367456c7 | From | Tim Chen <> | Date | Fri, 20 Apr 2012 09:40:42 -0700 |
| |
On Fri, 2012-04-20 at 16:00 +0200, Peter Zijlstra wrote:
> > Ok, so I can't reproduce this on my WSM-EP.. even !PREEMPT kernels are > consistent with hackbench times with or without that patch. > > Can you still send your full .config? Also, do you have cpu-cgroup muck > enabled and are you using that systemd shite? > > What does the below patch (on top of the previous) do? >
There is a slight 10% to 15% improvement with the patch. However, the change in performance difficult to quantify precisely as the hackbench runtime has large variations (up to 50%) on our Sandy Bridge EP server machines since commit 367456c7.
We also do not see regression for hackbench on WSM-EP, but only on machines with Sandy-Bridge EP (2 socket, 8 cores/socket, HT enabled). We are not running hackbench in cgroup for this test.
The Sandy Bridge EP machines installed has FC16, so I think it uses systemd. The machine on our WSM EP has FC15, which also has systemd.
I'm sending you the .config in a separate mail. Thanks for taking a look. The .config aligns closely with standard Fedora settings.
Tim
| |