Statistics (for server)

[Peter Polkinghorne]

Just recently I upgraded to Samba 2.0.6 from 1.9.18p10.  I noticed server
loading increasing and identified one candidate for causing this -
	change notify timeout
changing this from default of 60 seconds to 120 and then 240 produced a
significant reduction in load - see my message in samba mailing list - subject:

change notify timeout Performance issue

This made me realise - it is hard to know what your Samba server is doing.
I used truss (to trace system calls), smb.log (with normally debug level 1
but raised on selected processes to 3) and experiments with clients.
Just raising logs levels submerges you under huge volumes of data.

I have a little perl script that condenses smbstatus output into a single line 
to record number of connections, shares, deny mode locks and oplocks plus 
shared memory usage.  This is recorded at 10 minute intervals for all our 
servers.  This does allow me to say number of clients was much the same 
between days.

But I have no measure of how active such clients are.  But what would be good 
measures?

I can think of Data transfered, packets exchanged and may be counts of 
operations by type - but that might be too large for SMB (cf nfsstat).

These could be held or dumped per share? or connection? (on closing).

I am happy to try modifying the code but wonder if this a good direction to go 
in.  My motivation is that there still seems to be some extra loading on the 
servers above that provided by "change notify" functionality.

Also the issue of sizeing of servers - if we knew what types of clients were 
doing, we might be able better to predict server requirements.

-- 
-----------------------------------------------------------------------------
| Peter Polkinghorne, Computer Centre, Brunel University, Uxbridge, UB8 3PH,|
| Peter.Polkinghorne at brunel.ac.uk   +44 1895 274000 x2561       UK          |
-----------------------------------------------------------------------------