On Thu, Oct 04, 2012 at 11:42:45PM +0200, Gregor Best wrote:
> Hi people,
> 
> after a (very) long time of silence on this, here's another go at it. This 
> time,
> I basically started from scratch and used a bit of code by Christiano 
> Haesberth
> which had been posted to tech@ a while ago to detect CPU topology on amd64 and
> take that into account when moving processes between CPUs.
> 
> This version has one single queue per CPU, getting rid of a) the one single 
> system
> wide runqueue and b) the queue for expired processes. This simplifies things 
> a bit
> and performs just as well as my previous versions (the only difference is the 
> order
> in which expired procs get selected for running on a CPU). One advantage is 
> that
> process selection is in O(log n) of the number of processes on the CPU and 
> depends
> neither on the total number of processes nor the number of expired processes 
> in
> the runqueue.
> 
> The factors for the cost of moving a process between hardware threads, cpu 
> dies
> and cpu packages are guesses now, I think those will have to be tuned further.
> Sadly, I haven't had access to a multiprocessor machine with a more diverse
> architecture than "a bunch of cores on the same die".
> 
> I tested this on some more machines than before; a Core i5, an i7 and my Core 
> 2
> Duo and on all machines (perceived) interactivity was improved. The simplest
> benchmark I used was playing back a 1080p version of Big Buck Bunny with 
> mplayer.
> All machines I tested on had Intel graphics cards, one GM965 (on the 
> Core2Duo) and
> the others were Sandy Bridge devices. On all of them, playback was smoother 
> with
> the i7 being most visible. With the default scheduler, watching the movie was 
> a
> big pain due to heavy frame-dropping, with my patch, the movie was watchable 
> (with
> frame dropping only (barely) noticable in scenes with much movement).
> 
> As before, I'm looking forward to anything you have to comment, especially 
> cool
> benchmark ideas or the like.

With your patch, html5 videos on firefox are smoother, even the playback
is better when I try firefox + dpb 2 jobs vs only firefox without your
patch installed. I use i3wm, so I can't comment about the performance of
a full desktop environment.

My hardware: AMD A4-3400, 8GB, ATI Radeon HD 4350.

Please, still working on this area :)

-- 
Juan Francisco Cantero Hurtado http://juanfra.info

Reply via email to