Stephen,

In the end the question is one of expectation and view point.
The statement in the Admin Guide is that AFS services can be installed
on multi-homed machines.  It starts off by explaining that "AFS support
for multihomed file servers is largely automatic" and why that is the
case.  It then says "On database server machines, the NetInfo and
NetRestrict files determine which interfaces the Ubik database ..."

Of course my reading of this document is colored by the fact that I have
read the source code and know the wire protocol and the database
formats.  As such I know that the only piece of the AFS protocol suite
that is multihomed enabled are the UUID enabled VL_GetEntryByNameU and
VL_GetAddrsU RPCs on the VL service and the cache manager's management
of file server addresses to the file server UUID bindings.

DB servers do not have UUIDs but more importantly the UBIK database
replication protocol has no support for them or any other mechanism that
could associate more than one IP address with a single service.  From
the perspective of UBIK and all of the DB server clients, each IP
address is a distinct server.  Since the UBIK master election algorithm
relies on the ability to distinguish voting servers and ensure that each
(except for the one with the lowest IP address) has only one vote, it is
necessary to use NetRestrict, NetInfo and perhaps clones to constrain
which interfaces are visible and what they can be used for.

Can AFS DB services be deployed on multihomed systems? Yes.

Are AFS DB services multihomed capable? No.

There are two reasons for using bonded interfaces with a single IP address:

 1. Higher maximum throughput

 2. Resiliency to network interface or switch failure

There is no need to introduce the complexity of multihomed configuration
for AFS into the mix.

The lack of multihomed support extends beyond the DB servers.  The vos
commands which in theory could take advantage of knowledge of which
addresses are assigned to which file server / volume server do not do
so.  In AFS 3.4 when the UUID enabled RPCs were added to the VL_ service
not all of the required functionality was implemented.  In AFS 3.5 the
unfinished work was removed.

You can call it what you will but as far as I am concerned AFS does not
behave on multihomed systems the way one expects that it should.

Jeffrey Altman

On 9/24/2013 7:04 PM, [email protected] wrote:
> I know I've played around with multihomed dbservers before, but it was
> long ago, a different use case, on a different platform, and I had
> problems then! :-\
> 
> [edit] I just checked rt.central.org and found my ticket. No update in 9
> years. <https://rt.central.org/rt/Ticket/Display.html?id=15640>
> 
> Anyway, I don't think I'm mis-reading the Admin Guide (but I concede
> that what it says may be wrong!)
> 
> In any event, the QuickStartUnix guide also says they can be multihomed:
> "It is best to maintain a one-to-one mapping between hostnames and IP
> addresses on a multihomed database server machine..."
> 
> But I know all of the openafs docs are often out of date or downright
> wrong, which is why I asked on the list.
> 
> As far as you know, is the "it doesn't work" definitive?
> 
> Cheers,
> Stephen
> 
> On Tue, 24 Sep 2013, Jeffrey Altman wrote:
> 
>> Stephen,
>>
>> I believe you are misreading the Admin Guide.  It indicates that File
>> Servers can be multi-homed and that they self-register.  If you have DB
>> servers that are multi-homed you need to pick a single network interface
>> for them to use
>>
>> If given the choice of interface bonding and multi-homed, choose bonding.
>>
>> Jeffrey Altman
>>
>> On 9/24/2013 6:01 PM, [email protected] wrote:
>>> I have two general questions about multihomed AFS servers to pose to
>>> list.
>>>
>>>
>>> - Is there any obvious reason to choose multihoming of a fileserver
>>> rather than link aggregation, assuming both are supported in a given
>>> environment?
>>>
>>>
>>> - The Admin Guide on docs.openafs.org indicates that multihomed DB
>>> servers work.[1]
>>>
>>> Is this true in an environment where both IPs are on the same (campus)
>>> network and they may share the same default route?
>>>
>>> Specifically, I'm curious about linux dbservers; clients may contact the
>>> server via IP A but get an answer from IP B, if it is the default
>>> interface. Will the clients care? What if the IP replying isn't in the
>>> client's CellServDB (or DNS) at all? Does the answer change depending on
>>> client version?
>>>
>>> [1] <http://docs.openafs.org/AdminGuide/index.html#HDRWQ138.html>
>>>
>>> Cheers,
>>> Stephen
>>>
>>> _______________________________________________
>>> OpenAFS-info mailing list
>>> [email protected]
>>> https://lists.openafs.org/mailman/listinfo/openafs-info
>>
>>
> _______________________________________________
> OpenAFS-info mailing list
> [email protected]
> https://lists.openafs.org/mailman/listinfo/openafs-info

Attachment: smime.p7s
Description: S/MIME Cryptographic Signature

Reply via email to