NetBench problem

[James Peach]

On 18/04/2007, at 5:08 AM, jiri sasek - Sun Microsystems - Prague  
Czech Republic wrote:

> Hi all
>
> We have the following problem:
>
> - QA team using netbench raised a problem to having a >20%  
> performance regression on samba shares between the 3.0.21c and  
> 3.0.24 versions of samba.

Try comparing packet traces of 3.0.21c and 3.0.24.

> - we tried to reproduce it by smbtorture using the Tridge's  
> client.txt and we obtained the performance diffs between the  
> 3.0.0-3.0.9-3.0.10-3.0.21c-3.0.24 "flat" (diffs was no more than 1-2%)

yes, unfortunately this is not always predictive of actual windows  
performance.

> - we trying to find a root cause of this different results. We found  
> the client.txt was obtained snooping the traffic of the NetBench of  
> the version ~=3.
>
> Questions:
>
> - is there anybody who has the client.txt derived from the NetBench  
> v7 traffic or who is experienced with the creating it?

I believe that samba4 has a backend that is capable of creating  
workload files, though I've never experimented with it.

> - is anybody know the differences which can cause such performance  
> drop?

IIRC the default for "msdfs root" changed to "yes" around this time.  
We found that enabling this causes windows clients to do a lot more  
requests, with a corresponding drop in (netbench) performance.

--
James Peach | jpeach at samba.org