Re: Source for Cr overhead

On 4/18/06, Brian Paul <[email protected]> wrote:
>
> Niraj Tolia wrote:
> > Hi Brian,
> >
> > We are using fairly recent Cr CVS code for QuakeViz (an earthquake
> > modeling app from http://www.cs.cmu.edu/~quake/>. Our testing right
> > should not be stressing a system as we rotate an small model in a
> > 640x480 window. However, we see a very significant difference in the
> > nubmer of frames/sec.
> >
> > Attached is a graph of a particular interactive benchmark.
>
> I don't see an attachment.


My apologies. It seems I forgot to attach the PDF. Attached now.

> Thick is
> > simply running the model on NativeGL (Linux) and measuring the number of
> > glxSwapBuffer() calls. Chromium loopback is using the Cr renderer on the
> > same machine as the application. Do you know why there is suck a stark
> > difference in performance? The overhead of using TCP + GL
> > Marshalling/Unmarshalling? Any thoughts would be helpful.
>
> Going over the network is always a lot slower than the local graphics
> card.  How much slower depends on various factors like the network
> speed itself (no surprise there).
>
> Using display lists is one potential way to increase performance since
> you can draw quite a bit with a few, small glCallList commands.
>
> Another simple thing you can do is run 'top' on the hosts to see if
> any of them are CPU limited.


I did not think this was a problem as the machines are 3.6 GHz dual-threaded
Pentium IVs but I will look into this.

Thanks,
Niraj

> PS I kept this email off list for now. However, if you feel it would be
> > helpful to others, feel free to cc the list on the reply.
>
> I'm cc'ing the Chromium dev list in case anyone else has anything to add.
>
> -Brian
>



--
http://www.cs.cmu.edu/~ntolia