Subject: Performane comparisons, 1.3.3-current
To: None <tech-perform@netbsd.org>
From: Frank van der Linden <fvdl@wasabisystems.com>
List: tech-perform
Date: 02/22/2001 19:48:42
A while ago, Thor Simon came up with the idea to run benchmarks
on different NetBSD versions on the same hardware, to and compare
them to see where and when performance was lost or gained.

I've done such a test, with NetBSD 1.3.3, 1.4.3, 1.5 and 1.5Q.

The results are at ftp://ftp.netbsd.org/pub/NetBSD/misc/fvdl/bench/

The tests were done using hbench. See the README in that directory,
and the hbench source code for more info. These are simply
the 'summary' files.

Highlights: a couple of things improved: lat_syscall (due to
a recent optimisation effort by Charles Hannum), and lat_mmap
(because of Chuck Silvers' UBC work).

However, several others got quite a bit worse over time. For example,
lat_connect (big increase between 1.3.3 and 1.4.3, so that's not due
to IPv6 integration), and lat_fslayer.

I've also done some comparisons on the i386 mp branch; for the 
none-MULTIPROCESSOR case the numbers are the same (as was to be
expected), but for the MULTIPROCESSOR case the locking adds quite
a bit of delay to each system call (null call went from about 2us
to 6 us, for example). That's a seperate issue though, since
we have no previous history with that, comparisons with other
OSs make more sense there. I have done single CPU comparisons
with other OSs, I might post those at a later time.

- Frank

-- 
Frank van der Linden                           fvdl@wasabisystems.com
======================================================================
Quality NetBSD CDs, Support & Service.   http://www.wasabisystems.com/