>>>>> "Daniel" == Daniel Bromberg <[EMAIL PROTECTED]> writes:
Daniel> Whoops, sorry about the previous blank message.
Daniel> The JPL cell administrators are interested in analyzing AFS/DFS
Daniel> performance and how it can be optimized. A possible concern is that
Daniel> the current network infrastructure might eventually be saturated by
Daniel> AFS traffic once several thousand users call AFS their home. Major
Daniel> questions are:
Daniel> 1) How does AFS performance compare with DFS under various comparable
Daniel> loading conditions? (i.e. # volumes and # of accesses per volume)
Some of this information was presented at Decorum this year. Look for a
paper by David Stepheson and Steve Berman, Steven. From those Proceedings.
I should also mention that there may be some more of this type of perf
info in an upcoming IBM publication.
Daniel> 2) Is there a good way to load-balance servers by having an "AFS
Daniel> rover" track usage and automatically move volumes around? Should such a
Daniel> tool be developed?
Yes, there is a tool called Balance that does this. Check the CMU Andrew
ftp site or their Web page.
Daniel> 3) Is the above concern a valid one? Has any site run into any
Daniel> bottlenecks at all, network or otherwise, using AFS or DFS? What's
Daniel> been done about it?
This is always a concern. It's also what separates good deployments
from bad ones. When you're talking big scale, you can't blindly ignore
these issues. Unfortunately, the number of dependencies are such
that it is not easy to come up with a "generic" formula for a design.
(Anyone who claims they have this is probably blowing smoke!)
Daniel> 4) What other AFS/DFS metrics and optimizing tools would be useful, or
Daniel> are already developed?
Yes, Bob Oesterlin and company from the IBM Rochester Site have some
customized versions of RxDebug and scout which do these things.
There are some more specialized tools, if you want to dive in deeper.
Xstat, modified tcpdump's, etc.... It really depends on what you're trying
to optimize. (Generic client response, Response of particular application
using AFS/DFS for file storage, .....)
Also, something like Intellisoft's DCE Sleuth may help a lot in analyzing
DCE RPC performance and allow you to optimize DFS traffic.
Daniel> 5) Are there any actual studies or research papers on any of these topics?
Yes. I would suggest getting copies of the Decorum and OSF DCE Developers
Conferences from the last few years. That's the best place to go
information grazing.
Daniel> 6) What are good sources to find out about all of the above?
I would suggest looking at CMU's Andrew Web page. Also, try big AFS/DCE sites
like U of Mich/CITI, MIT, Stanford, U of NC, Cornell Super Computing Center.
Daniel> 7) Does anyone actually care?
Yes, I do, as do many of my peers. You posted to the right place.
Daniel> Thanks much for any help or pointers! Anticipatingly,
Daniel> Daniel Bromberg, Co-op
Daniel> [EMAIL PROTECTED]
Daniel> M/S 126-130 (818) 393-3872
Daniel> Jet Propulsion Laboratory
Daniel> 4800 Oak Grove Dr.
Daniel> Pasadena, CA 91109
One last thing. Remember people will often attempt to improve performance
at the expense of reliability. Here at IBM Austin, we try to get a handle
on monitoring/detecting faults and failures first. I've seen many
cases where this concept has been ignored and the cure ended up being
far worse than the disease.
--
Chris Cowan
ISSC (DCE/DSM Architecture)
-------------------------------------------------------------------------
Phone: 512-823-0113 FAX: 512-823-0727
--
"Writing about music is like dancing about architecture."
Thelonious Monk
