Subject: nfs caching
To: None <tech-kern@NetBSD.ORG>
From: Ron G. Minnich <firstname.lastname@example.org>
Date: 02/26/1997 11:11:08
chuck cranor forwarded me a note via tech-kern about an nfs file caching
question. The writer was proposing caching in the MV. He thought I ought
to mention the autocacher I did in '89, which was written up in ( in
think) '91 winter usenix.
Anyways the autocacher ran like an automounter, save when you read files
they were copied to local disk and cached from there. I like the way it
works better than the solaris cfs; you can read the paper to see why
(it's on my web page: see below URL). In particular in solaris CFS since it
is a block cache there's lots of concerns re integrity, so that for
example you need to 'cfs_fsck' the cache when you come up. CFS is just too
complicated for my taste. I note that I haven't found that many people
actually using it: they seem to be scared of it for some reason.
I think I prefer whole file caching to to caching in the vm as the earlier
post proposed. Statistics I accumulated over ~100 machines show that if
you have, e.g. a 1 GByte /usr/local, for any given user you only need to
cache 5-10% of that (this is the high end), even using whole-file caching.
That's $10 of disk.
Ensuring that up-to-date files are being accessed is trivial, of course.
I had a version of autocacher that was implemented as patches to amd, if
anyone wants it. It will need work as the code is so old.
Ron Minnich |"Failure is not an option" -- Gene Kranz
email@example.com | -- except, of course, on Microsoft products