Samba performance tuning & benchmarks

[Jim Morris]

Hi all.  Many KUDOS to the Samba development team!

I am trying to interpret some performance testing results, and could
use some hela, and any possible Samba tuning suggestions.  Oplocks are
turned on, BTW, which has caused the Samba performance results to come
WAY up compared to with prior Samba releases.

The following test results consist of the time to perform database
queries using a Borland Delphi 32-bit application.  The database is
approximately 10MB in size, and is stored in Dbase format.  The two
server systems in the test are a Samba 1.9.18p2 (Linux 2.0.33) system,
and an NT Server 4.0 system.

It is interesting to note that although initial queries are slower
when accessing files on the Samba server, additional queries from the
same Win95 client are 10X faster than when accessing the same database
on the NT server.

Oh - and one last thing. The server hardware is comparable in both
cases, with the exception that the NT server has a P166, whereas the
Samba/Linux server has a slower P100 processor.  RAM and disk I/O
performance characteristics are comparable.  Although processor speed
could explain the discrepancy, I find it hard to believe, considering
that subsequent database queries are faster on the Samba server.

> Yes.  It seems slower to do the initial query on Darkstar, but
> subsequent queries are much faster.  NT seems to be faster on the
> initial query but each query after that takes almost as long as the
> first one.
>
> To give you an idea:
> 
>                            Darkstar         Deathstar
                             (Samba)          (NT 4.0) 
                             (P100)           (P166)
> Initial Open Calls query   30 secs.         17 secs.
> Open for Barnie's           1 sec.          10 secs.
> All calls for store 312     1 sec.          12 secs.


Again, thanks for any advice or explanation as to why the initial
database query takes so long.

/-------------------------------\
| Jim Morris  |  jim at morris.net |
\-------------------------------/