In a former life I was the AFS manager for a pretty big AFS site (Fermilab) It's a two-edged sword. Don't really even try to think of it in the same sense as NFS. The small web cluster you describe is sort of worst-case scenario for AFS - NFS will undoubtedly do much better for you. Now if it was a BIG web cluster - maybe The AFS cache is a big win if you have a lot of client systems, and you have slow (maybe world-wide) networks. And you want a consistent view of a file system from EVERYWHERE. This is really a wonderful part of life - you want to see what your colleague in Russia is up to? Go to /afs/somerussionuniversity/igor - and there you are. No mounting, no understanding where his files are. Also a big win if you typically use a small number of files over and over - they're cached right on your local system. AFS is also great because you can replicate volumes. So say you have - um - say the equivalent of a /usr/local/bin where you have binaries. You can set up AFS with replicated servers around your network holding these read-only files. Auto load balancing. Auto failover. Automatic connecting to the RIGHT BINARIES FOR YOUR PLATFORM. Also it's great administratively - I used to run a program at night that shuffled everyone's home volumes to load and space balance. Nobody even ever knew unless they were trying to write to a file at 3 AM - then it would hang till the move was over - But if you dont' NEED those things, and aren't willing to deal with the Kerberos overhead, and installing a whole new set of RPC's on every system etc. it's probably not for you. Basically its great if you have a lot of clients and a consistent architecture and enough data to make it worthwhile. But it's really not at all a drop-in replacement for NFS. It's a whole different animal. Some days I miss it. Some days I'm glad I don't have to deal with it. As for performance - don't think of it from a user perspective. It has some - um - issues from a normal user viewpoint . It will never seem as snappy as NFS - at least assuming fat network pipes. BUT - it will load your servers a good deal less in many environments, and load your network a whole lot less as well - but as usual, this depends. For example, the idiots who originally wanted it at Fermilab insisted on putting it in in the worst possible places first, considering its intrinsic behavior. It's really no good if what you're doing all day is reading 2 Gb files of experiment data. It's great for running things like software. Not at all worth considering unless you've got some problems that it solves - just like any other tool. And there's considerable management overhead up front. I dunno if that helps any? nate at refried.org wrote: > On Fri, May 03, 2002 at 12:31:35PM -0500, Bob Tanner wrote: > >>Quoting nate at refried.org (nate at refried.org): >> >>>AFS >>> - Client keeps a cache of 50 - 100 MB >> >>How about performance? Would AFS make a better file server for >>something like a smallish web cluster then NFS? > > > Good question. I would guess that the initial setup of an AFS cell > stops most people from answering such a question. I did a little > searching and couldn't find any performance numbers. I did find a few > interesting links: > > IBM Redbook on using AFS for web content management > http://www.redbooks.ibm.com/pubs/pdfs/redbooks/sg245857.pdf > > Performance of Several HTTP Demons on an HP 735 Workstation > http://archive.ncsa.uiuc.edu/InformationServers/Performance/V1.4/report.html > > Doesn't Amy work at a site with an AFS cell? Perhaps she could run a > few tests. > > Nate > _______________________________________________ > Twin Cities Linux Users Group Mailing List - Minneapolis/St. Paul, Minnesota > http://www.mn-linux.org > tclug-list at mn-linux.org > https://mailman.mn-linux.org/mailman/listinfo/tclug-list