[OpenAFS] slave volumes

Jimmy Engelbrecht jimmy@e.kth.se
15 Jul 2002 13:40:02 +0200


"klaas hagemann" <klaas@northsailor.de> writes:

> as far as i understood afs, i can create severeal read-only replicas of a
> volume. How is the decision made, which of these replicas is taken?

In the case of openafs/transarc:

Your client will pick one of the sites by random, sites with a lower
"serverpref" are more likely to be selected.

the serverpref is today calculated by pingtime and giving some bonus if
your are located on the same Class-C network.

Your client may keep accessing this replica untill it goes away.
(at least true for Transarc)


In the case of arla:

serverpref will be calculated by ping time and nothing else.  The client
will look at "serverpref" for EACH AFS-operation. Which gives true
loadbalancing for sites with almost equal pingtimes.

> We have a distributes WAN and want to avoid that the Client takes a volume
> which is behind a very thin connection. For availability reasons all the
> replicas should stay available but in an ideal situation only the volume
> with the widest or fastest connection should be taken first.

Has the faster connection also better pingtime ?
In this case youre lucky.

If the wider connection has better pingtime than the narrow one, it will
suck.

> Can someone explain me the algorithm used by OpenAFS?

There is no general algorithm, algorithm may be slighly different dependent
on the version you are using.

Arla 0.36 (or at least 0.37) will be using an completly different and more
adaptive algorithm.

/Jimmy