[vdr] Re: Introduction from newbie

[Emil Naepflein <Emil.Naepflein@philosys.de>]

On Thu, 11 Apr 2002 13:17:09 +0200 (CEST), Sergei Haller
<Sergei.Haller@math.uni-giessen.de> wrote:

> some years ago I did some tests for my net and 8k for both read and write
> packets was the optimum. I mostly use this value since then. but your
> milestone may vary.

I have development and tuning experience with nfs on very big servers
(not Linux) over all kind of media up to gigabit ethernet. My experience
is that normally not the media throughput is the limit, it is mostly the
protocol  implementation. For maximum performance you need NFS v3,
multi-threaded clients and servers doing intelligent read-ahead and
write-behind combined with large blocksizes. Going to 32 KB blocksize is
a first simple step.

I didn't have yet a look at the Linux implementation of NFS v3, but I
would assume that it should have at least a few MB/s throughput on a 100
Mbit Ethernet with multiple concurrent threads reading and writing. It
is much much harder to get 10 MB/s throughput with a single stream, but
this isn't necessary here. 

For measurements you should read large files with dd into /dev/null and
write large files with data generated by /dev/zero. This is similar to
the situation with vdr. Copying files will not give the right results.

Emil