Tue Sep 17 11:15:53 EDT 2002

Now that Clp is available on cvs or as a tarball I thought I should make
some comments on Clp.  You can run the Osi unitTest and it should solve all
netlib test set.  To make a more flexible executable do make unitTest in
Clp directory.  This creates an executable - clp which can be used in a
simple way e.g. clp mpsfile will solve using primal simplex.  clp mpsfile
-dualsimplex will solve using dual simplex.  clp - switches to reading
commands from stdin.  Obviously this interface needs improvement, but it is
a start. (minor note - the current tarball had scaling off by default, I
have changed to on in cvs).

It seems reasonably reliable and I am still working on stress tests.  If
anyone has a problem which fails, I would be interested to have  a copy of
the problem and the parameter settings.  I have tested it on LP's of > 1.5
million constraints, MIPs of >0.5 million constraints and in Branch and
Cut.

Performance:

The aim was to be reliable and not too slow.  For speed I am comparing to
OSL at present.  OSL has a presolve, which is not yet in Clp and it tries
to adjust settings and the algorithm to the problem so it is faster on
netlib test set.  I have also switched off presolve and forced primal and
dual in OSL to get a more even comparison.

On a 1.7 GHz Pentium 4 laptop with gcc (times in seconds):

                                        OSL                CLP
unitTest                         181                  461
Primal/ no presolve   342                 461
Dual/no presolve       353                 414

On an IBM Unix P690 using xlC I get for dual

Dual/no presolve       195                193

Also I had put in a trivial Branch and Bound code as required by
OsiSolverInterface but I am going to move it to a different directory and
decouple it from Clp and allow it to have cuts etc.  The aim is to provide
a very lightweight counterpart to Bcp.  I should have something up in a
week or so.

John Forrest