[Libdlo] [PATCH] udlfb: high-throughput urb pool

[akephart at akephart.org]

Bernie, 

On December 18, 2010 at 1:40 PM Bernie Thompson
<bernie.thompson at gmail.com> wrote:

[some deleted]
 
> Another parameter to adjust is the defio delay (defined in udlfb.h). I
> think it's currently set to 5 (nearly immediate scheduling) because I
> subjectively felt mouse lag at higher values.  But that means we're
> probably getting notification of only a few dirty pages at a time. And
> for the video playback scenario, almost certainly a longer
> accumulation of dirty pages would be more favorable.
>
 
[AK] I'll give this a look and see what kind of difference it makes.
 
> If you find a better value for that parameter (accounting for what's
> basically the opposite ends of the spectrum - mouse movement vs. full
> screen video playback), we can definitely update.
>
> If anyone can come up with a strategy for dynamically adjusting the
> value (e.g. triggered by thresholds on how many buffers are unconsumed
> out of our pool), that's definitely worth considering too.
>
> Best wishes,
> Bernie
 
You know, in all of this, It strikes me that the on-demand model should
solve the problem for both damage and defio models, since it retains the
performance of the nonblocking release while avoiding most of the pixel
tossing.
 
-andrew