[Samba] Samba Performance for Digital Recording

[Manuel Elgorriaga]

Hello,

We are doing digital recording over a 100 Mb/s switched network to a
dedicated Samba NAS-Filer and need to optimize our workflow and present
some performance data. I just want to be sure, to not overlook some
important aspect in my experimental setup. Let me explain in short what
is our infrastructure:

The Swiss Library for the Blind has started producing Digital Talking
Books (see www.daisy.org). Our recording studios are equipped with
Intel-PC's (Xeon 2 MHz, 1 GB RAM, 100 Mb NIC, SCSI-HD's) running
specialized recording software under WinXP. The recording settings are
22.05 Hz sampling rate, 16 bit and mono, resulting in a continuous data
stream of 353 Kbit/s (151 MB/h). The raw data of a produced digital book
may have a size of a few GB. Some PC's are used for pre- and postediting
of the recordings (over the network again), which generates less data
flow but may deal with many very little and also some quite big files.

The data goes through two dedicated Cisco-Switches for this
audio-subnet: a 100 Mbit first and then to a 1 Gbit backbone switch, to
which is connected our audio-NAS-Filer. This is a Xeon-Server with 2 GB
RAM, two Gbit NIC's and a 420 GB HW-SCSI-RAID-5, running Samba 2.0.7a on
the XFS-filesystem (instead of ext3 or Reiser) on Linux 2.4.16 (SuSE
8.1). Later transfers include moving these data to a digital archive
over the network.

Now I want to measure data transfer rates over this network (and CPU-,
IO- and memory use) at the NAS-Filer, with different workload coming
from the recording/editing-PC's. I'll do repeated single as well as
multiple parallel recording/editing sessions of about 10 min, and I want
to measure at intervals of max. 5 seconds (as time series), using the
sysstat (vmstat, iostat, mpstat) and tcpstat and netacct tools. My
questions are now:

- are these the correct tools for giving me the right values, which I
could further consolidate and analyze in some tablesheet  (I cannot buy
commercial ones, and I think SNMP/MRTG is not the right tool in this
case)?
- what will the bottlenecks probably be: Disk-IO at the RAID-system, the
filesystem itself, Samba?
- what are theoretical values for each part of this data transfer chain
(NIC's, Switches, CPU+RAM, RAID, Filesystem, Samba, ...), and where can
I get comparative values?
- any other hints/caveats, e.g. what is the penalty of putting the
interfaces into promiscuous mode?

Thank you very much in advance,
Manuel Elgorriaga Kunze

-----
Manuel Elgorriaga Kunze
Swiss Library for the Blind and Visually Impaired
Grubenstr. 12 / CH - 8045 Zurich
Phone +41 43 333 32 32 / Fax + 41 43 333 32 33