Re: X much slower in 2.0.24 than in 1.2.13

Linus Torvalds (torvalds@cs.helsinki.fi)
Sat, 2 Nov 1996 12:52:38 +0200 (EET)

On Fri, 1 Nov 1996, Steve VanDevender wrote:
>
> I noticed this same thing yesterday when benchmarking XFree86 3.2. It
> does look like it's potentially a kernel problem; there seems to be a
> much higher proportion of system (kernel) time to user time while X
> benchmarks like xbench and x11perf are running than seems reasonable,
> about 50/50 if xosview is believable. The benchmarks themselves seem to
> draw in a way similar to what would happen if they were being run with a
> high CPU load on the system, with jerkiness in the drawing.

[ Somebody else reported that the difference is between 2.0.23 and
2.0.24, so it looks like something specific to .24 ]

Ok, can you tell me your hardware? The 2.0.24 tulip driver has a problem with
the multicast code, for example, and that has been reported to result in some
strange behaviour.

(To fix the tulip code, search for a number like 0x08000 and add three zeroes
to the end: I think there should be only one case of that number in the file)

Another thing to check for: do you use gpm? There was a bugfix with VC
selection wait-queues in 2.0.24, and maybe that fix made gpm act
strangely. Maybe gpm _is_ adding CPU load to the system...

Also, Alan did a socket write-allocation change, which I haven't even tested
(I've never run 2.0.x since I upgraded to 2.1.x, and I haven't ported all of
the fixes from Alan into 2.1.x yet). Alan, have you checked the wmem changes
wrt performance?

Finally, the SMP TLB invalidate code changed, are you running on SMP?

Linus