Rens Troost <[EMAIL PROTECTED]> writes:
> client level does not buy you all that much. Concentrating the AFS
> cache at the workgroup level (share a sparc or rs6k box among 10-50
> users as their sambified afs gateway) makes good economic sense; you
> share the disk space for the cache among users with largely
> overlapping cache needs, and (more importantly) you get intermediate
> cacheing instead of end-node-only caching; a big win for people on a
> WAN.
I recall that CITI at the Univ of Michigan did some experiments to
study the benefits of intermediate caching (I believe it was called
iAFS). They gave the intermediate server about 8-10 Gig of disk cache
(ie, essentially an infinite cache); they were trying to avoid a
problem similar to going across a WAN -- they were going `across' to
an IBM mainframe.
I don't recall this exactly and probably have it wrong, but here it
is: one of their conclusions were that overall system utilization was
probably better if they divided that 8-10 Gig across the several
Macintoshes so that each Mac had about 15-20M of disk cache. Something
about one person's working-set purging all the others'. I think the
paper appeared either in Usenix 1992 or in one of the AFSUGs
proceedings.
> Again, not necessarily an advantage, depending on what problem you are
> trying to solve. If you have a lot of small offices at the end of
> narrow pipes, you win big only bringing new software across once.
I would rather do this by putting a volume replica across the WAN and
setting appropriate fileserver preferences; however AFS does limit one
to 13 replicas. Not very useful for data that changes more frequently,
but then my guess is so would be any solution that expects to maintain
data-consistency across a WAN while expecting data to change
frequently.
