Disable read/write caching to disk?

[Peter Jeremy]

On Fri, 2005-May-27 13:09:11 -0400, Charles Swiger wrote:
>Apple's Xsan clustering solution relies on a so-called "metadata  
>controller", which keeps track of locking and provides syncronization  
>and invalidation notification when the filesystem metadata changes.   
>SAN clients still read the actual file data directly via fibre  
>channel, but they also need IP-level connectivity via ethernet (an  
>unroutable LAN is fine) to the MDC.

This is roughly the same as AdvFS on the la[te]st version of Tru64
clustering.  Both Tru64 and Solaris _require_ a cluster-private LAN
(or equivalent - there's a PCI memory-to-memory card that can be
used as an alternative).

I suspect a master metadata manager and coherency controller with all
the other systems as clients is probably the optimal solution.  A
distributed MESI-style approach would look cleaner but be much more
effort to implement and I suspect the inter-node traffic requirements
would kill scalability.

>The MDC is not a single-point-of-failure since redundant MDC's can be  
>set up, but I believe if all MDCs go down,

Presumably there's MDC failover as part of the cluster management.
(At least that's the way that Solaris and Tru64 work).

Any UFS guru's want to comment on the [im]practicality of clustering UFS?

-- 
Peter Jeremy