|
>
> > What about NFS?
>
> NFS is fine to maintain state on LVS backends.
>
> > I saw awhile back that the LVS FAQ said "NFS for clusters is
> > dead."
>
> not quite; a quote from the beowulf mailing list, referring to
> the use of NFS in beowulfs says "NFS for clusters is dead".
> I agree with this, but a beowulf is not an LVS. The reasons
> why NFS does not work well in a beowulf (file propagation
> scales with O(n) rather than something better like O(logn),
> and is network intensive), should be used as cautions when
> designing an LVS. In an LVS, if NFS is propagating read-mostly
> files to the realservers where they will be cached and where
> writes are not particularly time sensitive, then network traffic
> will be much less and NFS is fine.
>
I'm not sure you can always say network traffic is the reason why NFS write
performance is slower than local I/O (I know you know this but I thought I
should say it anyway). In fact, a 100 Mbps pipe to the switch works just
fine for a 10+ node cluster with 3,000-4,000 (NAS) transactions per second
(the NAS server is connected to the Ethernet backbone at 1 Gbps). My
experience is that slow NFS write performance has more to do with excessive
GETATTR calls that are triggered each time a record is locked.
> The other problem with NFS is that you have to handle the connection
> to realservers which have just died.
>
As far as the NFS-side of things goes the NSM implementation in Linux (as an
NFS client) is good enough in production (in my experience) to prevent
deadlocks and stale or hung locks. The client connection dies, but at least
when they reconnect they can get the record lock they want back (talking
about the data not the session information).
> > But my book, and the cluster I built is based on NFS as the CFS. I'm
> > not a DB guy, but as I understand it Oracle's 9i RAC product is based on
> > running Oracle servers on NFS clients with a NAS backend. Their strategy
is
> > to get all of their customers to move to the RAC line.
>
> Oracle's goal is to make money for Oracle. I wouldn't regard
> anything Oracle did at the technical level as a pointer to good
> techical practice. It might be, but I would have to go elsewhere
> to find out.
>
True enough, but another place to look is at what is happening with data
storage especially NAS.
> > Why the harsh treatment of NFS in the FAQ?
>
> I just say that you shouldn't use NFS where it doesn't work.
>
Yeah, I'm thinking more of transaction-based computing and not scientific
applications.
Just my 2 cents.
--Karl
|
