[OpenAFS] Re: about openafs , fileserver process , performance

Andrew Deason adeason@sinenomine.net
Thu, 20 Feb 2014 12:20:32 -0600

On Thu, 20 Feb 2014 14:33:15 +0100
nicolas prochazka <prochazka.nicolas@gmail.com> wrote:

> I do some benchmark ,
> cat /afs/mycell.com/bigfile > /dev/null   on openafs clients
> and i see a bottleneck of 25MBytes/s  on server side ( fileserver process) .
> My ethernet link are 1Gbit/s
> How can I suppress this bottleneck ?

Do you know what the RTT is between the server? (just 'ping' the server
for an approximation)

The maximum theoretical throughput you'll get from openafs right now is
somewhere around ((32 * 1400) / RTT) bytes per second (due to some
limitations in the transport mechanism). That means if your RTT is on
the order of 10ms, the maximum throughput you'll see is is 4 or 5 MiB/s.
If your RTT is around 1ms, you'll see a maximum of around 40 or 50
MiB/s. That is the theoretical max, so the actual throughput you'll see
will be less.

There are other potential bottlenecks, of course, but that is often the
biggest factor when someone is surprised at the low transfer rates that
they see. There are some efforts underway to alleviate those
restrictions, but you can't do much about it just by "tuning" clients or
servers etc.

Andrew Deason