Re: Quake strangeness

Martin Ostermann (
06 Nov 1996 12:05:34 +0100

Linus Torvalds <torvalds@cs.Helsinki.FI> writes:
> On Tue, 5 Nov 1996, Bernd Bernie Meyer wrote:
> >
> > Then, I got even sillier and started _another_ copy of Quake on the
> > ALPHA's local X-Server. And here the strange thing happened:
> >
> > While the copy that had to send all its data to the Pentium reached
> > 16fps, the local copy only reached about 6fps. According to top, they
> > both got equal CPU percentage; Also according to top (and vmstat), app.
> > half of the CPU-time was spent in system mode.
> >
> > I can't seem to make sense of this --- why is the local copy so much
> > slower than the remote one? I tried running two local copies, and
> > they both get app. 20fps, which is as expected. I tried running two
> > local plus one remote copy, and the local ones became SLOOOOW
> again.

So, I understood this: Two copies of Quake on the Alpha, displaying
locally, get each 20fps. Of two copies of Quake on the Alpha, on
displaying locally, the other remotely, the first produces 6fps, the
later 16fps. Total frame rate thus is 22, vs 40 before.

So sending 16fps over lowers overall performance by 50%, vs displaying
it locally. This is the 50% spend on system time, which seems
a lot to me.

> > Is this a scheduler thing, where the remote Quake process gets scheduled
> > all the time due to network interrupts? But if so, why does top show all
> > copies getting equal CPU time?
> It's probably simply because the time-consuming part of quake is not so much
> the actual drawing as the _calculations_ it is doing.

Well actually the time consuming part seems to be to send the data over the
net. Probably sending anything at that rate (16fps*320*200 is
1MByte/s) costs 50% CPU time of anything calculated at the same time.

> And with the remote copy, the load is a lot lower. With a local copy of
> quake, the local load is pretty high, so the local quake binary has
> trouble getting CPU for calculations, while the remote quake obviously
> doesn't have that kind of performance problems..

Maybe the time spend in system mode should be isn't accounted
correctly to the process causing it?


Martin Ostermann                |
Communication Networks          |
Aachen University of Technology | phoneto:++49/241/807917
Germany                         | faxto:++49/241/8890378

To unsubscribe: send e-mail to with
'unsubscribe' as the subject.  Do not send it to

Feedback | Store | News | Support | Product Errata | About Us | Linux Info | Search | JumpWords
No Frames | Show Frames

Copyright © 1995-1997 Red Hat Software. Legal notices