dbench scalability testing

[Anton Blanchard]

>   I work for the OSDL and we are attempting to script some dbench
>   run-throughs to help us test the scalability of some of the samba/Linux
>   kernel code.

Cool.

>   The problem is, using a single kernel image (2.2.18) on the same hardware
>   without any other system activity I can't get reliable results.

At this point, testing on 2.4 would be better. Most of us are working there
and not on 2.2.

>   I have taken precautions to exclude caching issues from the list of
>   possible problems, but for some reason I still can't get the results to be
>   reliable.
> 
>   I have posted the results at the following location, if anyone can offer
>   some help, I would appreciate it.  Right now, booting and running the
>   benchmark gives different results than running the benchmark after 8 or so
>   hours of uptime.
> 
> http://63.68.113.130/~smurf/Samba%20dbench%20Results/sequence-2/

I have been able to get reliable results with recent 2.4 kernels. For
benchmarking purposes, increasing the async and sync flush points helps:

/proc/sys/vm/bdflush:

before:
30	64	64	256	5000	30000	60	0	0
after:
90	64	64	256	5000	30000	95	0	0

Netbench is such a bogus benchmark but we are stuck with it and we can
optimise for it nicely :)

>   If you want more info on the OSDL, please see www.osdlab.org - we have 4 &
>   8 and (soon to be available) 16 processor boxes for use by the open source
>   community.  

Excellent, I've been doing samba/linux performance work but at this point
I only have a quad cpu machine. Can your network cards do zero copy? I
have a few samba patches including making samba use sendfile which could
help when you get around to doing netbench runs.

Anton