[OpenAFS] Using CPLEX for AFS server balancing

[Russ Allbery]

Hello folks,

At various AFS get-togethers over the years, I've mentioned the AFS server
balancing method that Neil Crellin developed here at Stanford that phrases
the balancing problem as a linear programming problem and then uses a
commercial linear programming optimizer to find a solution.  (Also known
as the *really thorough* balancing method.  *heh*)  A few people have
expressed interest out of morbid curiosity, so since I just finished
resurrecting this approach so that we could use it again even though our
cell has grown significantly, I took the time to write it up.

The overview and general discussion is at:

    <http://www.eyrie.org/~eagle/notes/afs/balance.html>

If you want to see the constraints and algorithm described there expressed
in mathematics, there is an accompanying mathematical paper at:

    <http://www.eyrie.org/~eagle/notes/afs/balance.pdf>

(A TeX version is also available.)  Finally, the software that we use to
prepare the AMPL input file and process the results is available at:

    <http://www.eyrie.org/~eagle/software/afs-balance/>

I would not really recommend that you use this approach unless you already
have a license for ILOG CPLEX or some other suitable linear programming
optimizer, since this probably isn't sufficiently better to warrant
purchasing anything.  If, like us, you just happened to have a license
sitting around, this approach will give you a rather thorough balance.

Note that the software (although not the technique) is dependant on the
AFS reporting database that I mentioned in my message to the list earlier
this week.

-- 
Russ Allbery (rra@stanford.edu)             <http://www.eyrie.org/~eagle/>