Subject: Re: NetBSD 1.5 on uVAX II (Questions)
To: None <>
From: Havard Eidnes <>
List: port-vax
Date: 01/01/2001 17:17:11
> > I don't think the size of the kernel matters all that much speed-
> > wise, unless you are severely memory restricted.
> OK, but I wanted to compare stripped identical kernels to see what
> bloat difference there was between the two.

I think that's a different goal than the one we were discussing.
I'm not saying that it's uninteresting to find the answer to this
question, but it should not be glommed together with finding the
figures for overhead of certain parts of the system.

> > What's otherwise the important thing is to run a kernel config
> > that's comparable (as few differences as possible) on the two
> > releases.
> Granted, but I wanted the gurus to send me their chosen config.
> If not, I will choose my chosen stripped config and run with that.

That's hard to do without knowing the exact details of your
hardware configuration.  I'd therefore suggest that you're the
best placed person to carry out this task.

> ``Who can name/write that config in 20 lines or less?''.....(:+}}.

In this case it's probably not the point; the point is to pick a
realistic test environment which doesn't have too many extraneous
differences between the configs used on 1.4.3 and 1.5.

An additional variation could be to compare the GENERIC kernels
from both releases, and see if that changes some of the figures
in a significant way.

> > I think I would recommend you run lmbench on the two systems,
> > and merge the results (give them different names so that the
> > results can be summarized in a single table for easy
> > comparison).
> OK.  Where is that, in the system or in the packages?

Um, package (surely you knew your way around NetBSD enough
already to know that? ;-) The source to the package was to be
found in pkgsrc/benchmarks/lmbench/ last time I had a look.

I'm not sure if it would make a difference in 1.5 whether the
program you run is compiled on 1.4.3 or 1.5; it's perhaps an
additional variation you can toss in.

> Well, the simple things I was doing was logging boot times,
> compile times, tcpip times, and that sort of thing.  I need
> to get better system call overhead numbers and that sort of
> thing to help find out where the bottlenecks are.  I will
> work on that.

lmbench can give you system call overhead for some selected parts
of the system, and (I think) will also give you some network
performance figures if you have a neighboring system to test
against (I've not actually done the latter myself).

> Query... I just downloded todays sources, and was wondering if I
> should compile those for the kernels or use the stock 1.5 suite?
> Anyone have any feeling on that, so as to provide the best info?

I'd suggest that for the moment you're best served by shooting at
a non-moving target, i.e. stick with 1.5.  You can always bring
in that additional variation later if you so desire, but I would
suggest that you produce the numbers for the stationary targets


- H=E5vard