Re: [SCHED] wrong priority calc - SIMPLE test case

[Con Kolivas <kernel@kolivas.org>]

[-- Attachment #1: Type: text/plain, Size: 2657 bytes --]

On Friday 13 January 2006 21:46, Paolo Ornati wrote:
> On Fri, 13 Jan 2006 12:13:11 +1100
>
> Con Kolivas <kernel@kolivas.org> wrote:
> > Can you try the following patch on 2.6.15 please? I'm interested in how
> > adversely this affects interactive performance as well as whether it
> > helps your test case.
>
> "./a.out 5000 & ./a.out 5237 & ./a.out 5331 &"
> "mount space/; sync; sleep 1; time dd if=space/bigfile of=/dev/null
> bs=1M count=256; umount space/"
>
>   PID USER      PR  NI  VIRT  RES  SHR S %CPU %MEM    TIME+  COMMAND
>  5445 paolo     16   0  2396  288  228 R 34.8  0.1   0:05.84 a.out
>  5446 paolo     15   0  2396  288  228 S 32.8  0.1   0:05.53 a.out
>  5444 paolo     16   0  2392  288  228 R 31.3  0.1   0:05.99 a.out
>  5443 paolo     16   0 10416 1104  848 R  0.2  0.2   0:00.01 top
>  5451 paolo     15   0  4948 1468  372 D  0.2  0.3   0:00.01 dd
>
> DD test takes ~20 s (instead of 8s).
>
> As you can see DD priority is now very good (15) but it still suffers
> because also my test programs get good priority (15/16).
>
>
> Things are BETTER on the real test case (transcode): this is because
> transcode usually gets priority 16 and "dd" gets 15... so dd is quite
> happy.

This seems a reasonable compromise. In your "test app" case you are using 
quirky code to reproduce the worst case scenario. Given that with your quirky 
setup you are using 3 cpu hogs (effectively) then slowing down dd from 8s to 
20s seems an appropriate slowdown (as opposed to the many minutes you were 
getting previously).

See my followup patches that I have posted following "[PATCH 0/5] sched - 
interactivity updates". The first 3 patches are what you tested. These 
patches are being put up for testing hopefully in -mm.

> BUT what is STRANGE is this: usually transcode is stuck to priority 16
> using about 88% of the CPU, but sometimes (don't know how to reproduce
> it) its priority grows up to 25 and then stay to 25.
>
> When transcode priority is 25 the DD test is obviously happy: in
> particular 2 things can happen (this is expected because I've observed
> this thing before):
>
> 1) priority of transcode stay to 25 (when the file transcode is
> reading from, through pipes, IS cached).
>
> 2) CPU usage and priority of transcode go down (the file transcode is
> reading from ISN'T cached and DD massive disk usage interferes with
> this reading). When DD finish trancode priority go back to 25.

I suspect this is entirely dependent on the balance between time spent reading 
on disk, waiting on pipe and so on.

Thanks for your test case and testing!

Cheers,
Con

[-- Attachment #2: Type: application/pgp-signature, Size: 189 bytes --]