[OpenAFS] Re: [OpenAFS-devel] Support for huge (really huge!) cell

[Ken Hornstein]

>On Sat, 6 Nov 2004, Ian Delahorne wrote:
>
>>> We plan to install a read-only AFS-DB-Server in every school too (makes
>>> at least 100). Will Ubik scale to this number of Servers?
>>
>> Ubik doesn't scale beyond three servers.
>
>Oh. There are (or at least were) cells with 5 Servers around.

I'm not sure where Ian got that number from; I personally don't believe
it, but I'd be glad to see proof that I'm wrong.

>And the other question ist: Do the read-only-DB-Servers take part in the
>election, when they can never become sync-site?

If you're talking about the "clone" database servers ... then the answer is
yes.  The only difference (in terms of the Ubik algorithm) is that if a
beacon request is received from a clone, nobody will vote yes for it.

>Our idea is, to use different CellServDB-Files in each school containing
>just the local RO-Servers and the tree Master-RW-Servers. Only the
>Master Servers will know all other Servers.

I'm not sure what happens if you need to talk to the master, but it's not
in your CellServDB.

>Sorry for my bad English. What I wanted to ask is: Does afsd use the
>same algorithm when it locates a Ubik-Server. In ubikclient.c (which is
>used by klog too, that's why klog went into this context) is a rather
>simple randomisation of the CellServ-List. Location of the Servers is
>not taken into account.

I believe that it's the same algorithm.  But the one issue is that afsd
will remember when a particular server is down; klog will not.

--Ken