>This latter configuration would be quite flexible, although it would be network
>traffic
>intensive: run the AFS server on machine1, using a "disk device" which connects
>to the HSM disk server on machine2. All traffic eventually going to the client on
>machine3.
> I say it is flexible because you could use any filesystem as the "disk server",
>just by accessing blocks and lseeking.
> Layering, ugly.
Well, as I understand your question ... MR-AFS works that way
already. The fileserver I/O layer has been replaced with a "remote
I/O" layer, and special auxilliary vnodes point to the actual
location of the data. Data access is funneled through the remote
I/O layer to "remioservers" running on the machines that store the
actual data. These residencies are filesystem independant - they
use the standard Unix system calls to access data, and as a result
they work on any Unix filesystem. Since all of the HSM systems I've
seen present a Unix filesystem as the interface, this works quite
well with MR-AFS.
--Ken
