More performance notes..

From: Dave Hudson <dave_at_nospam.org>
Date: Fri Jan 27 1995 - 06:39:21 PST

Hi All,

I thought I'd just mention where I'm up to with the kernel optimisations.
I've now got the context switch time as measured by perf1 on my DX/4 cached
system down to 51 us under gcc 1.42 (although 2.6.2 doesn't actually help
that much - sometimes it produces slower executing code on a 486 :-)).

I've also clobbered a couple of minor problems with the kernel debugger and
the kernel print routines.

David Koogler mailed me a port of IOzone which I've modified slightly (just
the timings, not the tests). I compared results on my system between the
standard 1.3.2 kernel and the one I now have. My 2.5" drive doesn't have
any write cache, and only 8k of read cache, so it's not what it could be,
but here are some fairly representative results (only difference is the
kernel - note k/s is 1000 bytes per sec, not 1024 bytes per sec):

v1.3.2 standard:

Test file size Record size Write rate Read rate
1 MB 512 98k/s 455k/s
1 MB 1024 99k/s 537k/s
1 MB 4096 101k/s 635k/s
1 MB 32768 109k/s 676k/s
1 MB 65536 109k/s 537k/s <- ???
8 MB 512 77k/s 551k/s
8 MB 1024 78k/s 696k/s
8 MB 4096 78k/s 847k/s
8 MB 32768 80k/s 887k/s
8 MB 65536 80k/s 687k/s < - ???

New faster kernel:

Test file size Record size Write rate Read rate
1 MB 512 101k/s 599k/s
1 MB 1024 103k/s 676k/s
1 MB 4096 103k/s 748k/s
1 MB 32768 109k/s 748k/s
1 MB 65536 109k/s 723k/s
8 MB 512 78k/s 776k/s
8 MB 1024 79k/s 911k/s
8 MB 4096 79k/s 1004k/s
8 MB 32768 80k/s 1010k/s
8 MB 65536 80k/s 1010k/s

To be fair, the 1.3.2 standard kernel has debugging turned on - with
debugging on in my kernel the results are somewhere on the positive side,
in-between - here are some results:

Test file size Record size Write rate Read rate
1 MB 512 101k/s 537k/s
1 MB 1024 102k/s 635k/s
1 MB 4096 104k/s 699k/s
1 MB 32768 110k/s 748k/s
1 MB 65536 113k/s 635k/s < - ???

Anyway, I know there's still room for improvement in the kernel, but I think
maybe finding ways of speeding up the dos and wd servers might be a good
idea now :-)

                                Regards,
                                Dave
Received on Fri Jan 27 06:00:59 1995

This archive was generated by hypermail 2.1.8 : Thu Sep 22 2005 - 15:12:17 PDT