[Intel-gfx] [PATCH] RFC drm/i915: Slaughter the thundering i915_wait_request herd

[Chris Wilson]

On Mon, Nov 02, 2015 at 11:26:29AM +0000, Gong, Zhipeng wrote:
> > -----Original Message-----
> > From: Chris Wilson [mailto:chris at chris-wilson.co.uk]
> > Sent: Monday, November 02, 2015 5:59 PM
> > To: Gong, Zhipeng
> > Cc: intel-gfx at lists.freedesktop.org; Rogozhkin, Dmitry V
> > Subject: Re: [PATCH] RFC drm/i915: Slaughter the thundering
> > i915_wait_request herd
> > 
> > On Mon, Nov 02, 2015 at 05:39:54AM +0000, Gong, Zhipeng wrote:
> > > Chris-
> > >
> > > The patch cannot be applied on the latest drm-intel-nightly directly.
> > > I modified it a little bit to make it applied.
> > > The patch can help much in HSW, but a little bit in BDW.
> > > The test is to transcode 26 streams, which creates 244 threads.
> > >
> > > CPU util      |	w/o patch  |   w/ patch
> > > ----------------------------------------------------------
> > > HSW async 1   |   102%     |     61%
> > > HSW async 5   |   114%     |     46%
> > > BDW async 1   |   116%     |     116%
> > > BDW async 5   |   111%     |     107%
> > 
> > Could I get the perf report for the kernel time? One aspect that I find hard to
> > believe is that it is not the execbuf/mutex-contention that is the ratelimiting
> > step.
> 
> Sure, what command would you like to run with "perf"?

Each of them :) I want to be sure that I know what's going on with bdw
(to check if my semaphores guess is correct), and comparing 1-vs-5
should help understand the contention points better.
As for the actual command, something like
  perf report -G -d '[kernel.vmlinux]' | head -5000
should do, though you may have to modify the DSO list to match if you
don't use a i915.ko builtin.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre