> To understand very well what' s happening to your system and to fix the
> bad behavior please now apply my lastest patch-ikd-arca (the latest is
> against 2.1.120). Maybe it apply clean also against 2.1.121. If you
> get only .orig files from the patch everything should be fine. You' ll
> find it at:
> http://e-mind.com/~andrea/patch_ikd_arca.shtml
> ftp://e-mind.com/pub/linux/patch-ikd-arca/
We will try this tomorrow and should be able to send you the results
(hopefully) tomorrow as well.
> The two traces will should everything we need to debug and fix your
> problem that I can' t reproduce here btw.
Ooops ! I tried to reproduce the problem (performance loss) with a
Digital Pentium 100 MHz (kernel 2.0.35) and a Pentium 200 MMX (kernel
2.0.35) without any success !
So I thought that the problem might be coming from the PPRO CPU
selection in the kernel configuration. We rebuild the kernel for our
dual PPro 200 MHz with Pentium selected as the CPU, we also removed
RTC (real time clock) and PCI bridge optimization. But again, we were
able to reproduce the performance loss.
So the problem really seems to be link somehow with Pentium PRO and
Pentium II processors !
> After you have logged the two traces you can use readprofile (you' ll find
> it at sunsite.unc.edu) to read the profiling info. So run readprofile
> --map (or something similar) System.map (sysmap of the ikd compiled
> kernel). Then send me the output compressed again. Then run `echo 1
> >/proc/readprofile`, run one istance of the proggy again and send the the
> second readprofile output. The profiling is not very important (it does
> not allow to understand what' s going wrong for you), is just that I am
> courious to see it ;-).
>
> Andrea[s] Arcangeli
Here is a sample output of a benchmark with our dual PPro 200 MHz (256
kb cache) running kernel 2.1.119 (built with PPro CPU selection):
After a clean boot:
init
1500 iterations in 14.997232 seconds with 0 overruns
1500 iterations in 15.000005 seconds with 1 overruns
1500 iterations in 14.999996 seconds with 0 overruns
1500 iterations in 15.000021 seconds with 0 overruns
1500 iterations in 24.786221 seconds with 146 overruns
1500 iterations in 69.839989 seconds with 705 overruns
1500 iterations in 17.843781 seconds with 33 overruns
1500 iterations in 14.999976 seconds with 1 overruns
1500 iterations in 15.000012 seconds with 0 overruns
Bench terminated.
85.430u 0.610s 3:25.98 41.7% 0+0k 0+0io 90pf+0w
When we launched around 50 other instances of prog1 we see that the
timing and overruns increase (which is completely normal).
After killing all instances of prog1 and restarting a new prog1
instance:
[radar2]~/Bench > ntt
init
1500 iterations in 23.543116 seconds with 2354 overruns
1500 iterations in 23.550882 seconds with 2355 overruns
1500 iterations in 23.562376 seconds with 2356 overruns
1500 iterations in 23.549635 seconds with 2354 overruns
1500 iterations in 23.529757 seconds with 2352 overruns
1500 iterations in 23.485237 seconds with 2347 overruns
1500 iterations in 23.557829 seconds with 2356 overruns
1500 iterations in 23.553188 seconds with 2355 overruns
1500 iterations in 23.554409 seconds with 2355 overruns
Bench terminated.
227.940u 1.180s 3:49.16 99.9% 0+0k 0+0io 89pf+0w
So our conclusions for the moment:
1- Kernels running on Pentium do not suffer from any measurable
performance loss.
2- Kernels running on PPro or PII suffer from performance loss
especially if the kernel is a 2.1.xxx
3- Kernels 2.0.xx are less affected
Regards,
Claude
-- Claude Gamache, CAE Electronique Ltee, 8585 Cote-de-Liesse Saint-Laurent, Quebec, Canada H4T 1G6 Email: cgamache@cae.ca Tel.: (514) 341-2000 x3194- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/faq.html