Re: Task switch timing tests?

David C. Niemi (niemidc@clark.net)
Wed, 19 Jul 1995 23:28:35 -0400 (EDT)


Hello Linus,

I did some quick benchmarks on my system (AMD DX4/100 PCI, 24MB RAM, 512
KB L2cache; mostly Elf system). I'll leave the full suite to Carlo Prelz.

The pipe-based context switching test (context1) showed very significant
improvements (on top of something like a 15% improvement since the 1.2.x
series). On my system it went from 7600 (1.2.10) to 9400 (1.2.11) to
14950 (1.2.11 + pre1.2.12 scheduler patch). This is about a 23%
improvement from 1.2.10 to 1.2.11, and a further 59% improvement to the
scheduler patch (for a total of about a 95% improvement).

This is about a 15% improvement over 1.2.8 + Matt Dillon's scheduler
patach, which means that 1.2.11+ the pre1.2.12 accomplishes about the same
or a tad better. Is the approach similar?

I did not see statistically significant differences in other benchmarks
(also true of Matt Dillon's patch) but I only tried a couple (-1% on
syscall, +2% on pipe throughput for 1.2.11+ vs. 1.2.10).

I see no problems with it either; in case you want a testimonial from
someone who has tried them for, oh, about 38 minutes.

DCN

On Tue, 18 Jul 1995, Linus Torvalds wrote:
> Hi,
> I just made 1.3.11 available on the normal ftp-sites (ftp.funet.fi and
> ftp.cs.helsinki.fi). I also have an experimental patch to go on top of
> that, and I'd like to know if it (a) works and (b) speeds up
> task-switching. Does anybody want to test this out (the byte pipe
> benchmarks should show the effects if any, and there are other purely
> context-switch sensitive benchmarks as well)?
>
> (I am actually using this patch on my own system, so it _should_ work ok
> for others too. I just don't have anything good to test context
> switches with).
>
> A plain 1.3.11 kernel will only make a rather small difference (the
> scheduling is cleaned up, but not much optimized), so the real
> difference should show up if you apply the accompanying patch on top of
> the 1.3.11 kernel (you get a kernel that says "1.3.12", but it isn't,
> really).
>
> Linus

----David C. Niemi---niemidc@clark.net---703-904-3596---Reston, VA USA----
We have a long and rich history, but no future save what we dare to dream.