Re: [g.mcgarry%ieee.org@localhost: More Microbenchmarks]

[Johnny Billquist <bqt%softjar.se@localhost>]

I'm curious. Why would it be of historical interest only, if we can 
produce numbers that show that the system worked much more efficiently 
in the past? (Ie. getting services done took less real time, using the 
same hardware, with similar loads.)

I found the numbers very interesting, and if it were me, I'd start 
looking into the places where performance have dropped significantly and 
try to understand why. Maybe there is a good reason, but then again 
maybe we're doing something stupid now that we shouldn't.

In any case, this fits well with the opinion expressed in port-vax a 
number of times that we actually seem to get a slower and slower system. 
But this have been disputed by others every time.

        Johnny

Thor Lancelot Simon wrote:
> On Tue, Jun 03, 2008 at 12:03:44PM +0100, Andrew Doran wrote:
> > ----- Forwarded message from Gregory McGarry <g.mcgarry%ieee.org@localhost> 
> > -----
> >
> > I've tried posting these results to tech-kern a few times without success.
> >  Feel free to forward these results along.
>
> I have comparable numbers for every NetBSD release from 1.2 through 1.6,
> but they're really just of historical interest at this point I think.
>
> I thought that there were a couple of very interesting regressions here.
> One that really caught my eye was that our local TCP bandwidth now exceeds
> our local AF_UNIX bandwidth -- that's just bizarre.
>
> The other was the huge increase in 0-byte file create/delete latency.
> But one thing to be careful with when using lmbench is that it tends to
> make sampling or arithmetic errors with very large or small numbers.  Is
> there any chance of rerunning this with hbench-OS?  It fixes some of those
> problems.
>
> Thor