Some additional questions about benchmarking CIFS servers

[David Lee]

On Sat, 22 Jun 2002, Richard Sharpe wrote:

> During some discussions yesterday with some folks, I realized that a 
> couple of additional deficiencies that NetBench has are:
> 
> 1. All the clients connect at the beginning of the benchmark, and do not 
> disconnect until the end of the benchmark. Connection handling is an 
> important aspect of a CIFS benchmark, and while you do not expect every 
> client to be disconnecting and reconnecting every five seconds, you do 
> expect a certain connection load, perhaps 5% of the overall clients would 
> connect and disconnect from the server twice during the benchmark.

I'll second that!  One of the main problems we had a year ago on our main
fileserver (Solaris, Sun 450, up to 1,000 simultaneous "smbd" connections) 
when we tried moving from Samba 2.0.x to 2.2.x (and then through the 2.2.x
series) was the avalanching of "smbds".  Contention on "connections.tdb"
seemed to be a contributory factor.  I understand that this was compounded
by oplock-related issues, but nevertheless, we were typically averaging
one operation per second on the database.  (Student classrooms spread
across campus;  "burst" activity as lectures finished;  etc.)  The machine
was already loaded with the routine activity of established smbds. 

So in our experience and environment, the connect/disconnect activity
seems to be a significant factor in loading a Samba server.

Hope that helps.

-- 

:  David Lee                                I.T. Service          :
:  Systems Programmer                       Computer Centre       :
:                                           University of Durham  :
:  http://www.dur.ac.uk/t.d.lee/            South Road            :
:                                           Durham                :
:  Phone: +44 191 374 2882                  U.K.                  :