[Coin-lpsolver] (no subject)

[Kish Shen]

Hi,

In one of the test I am running, I ran into a possible problem with Clp: 

the solver seems to be taking a very long time to solve the linear problem
l30, which is one of the problems used in Mittelmann's benchmarks. In his
webpage for the latest results (ftp://plato.asu.edu/pub/lpfree.html), he
lists the solving time for this problem as 80s, although it is not stated
what method was used to solve this problem.

I have setup Clp to use the dual simplex method via the OSI calls:

    lpd->lp->setHintParam(OsiDoDualInInitial, true, OsiHintDo, NULL);
    lpd->lp->setHintParam(OsiDoDualInResolve, true, OsiHintDo, NULL);

and so far, Clp have spent more than 80,000 seconds to try and solve this
problem, and as far as I can tell, it is still in phase 1, because the
objective value is not moving uniformly in one direction.

Both CPLEX and Xpress are able to solve this problem with dual simplex in
under 70 seconds, and SoPLEX in under 100 seconds. So far, for all other
problems, I have not seen such a big difference between the solving times
for Clp and the other solvers. Is this possibly a bug, or is it simply `bad
luck' because the pivot selections etc. happens to be very bad for this
problem for Clp?

Clp is able to solve this problem using primal Simplex in 208 seconds, so
perhaps the Mittlemann result is using primal simplex (his timings for the
other results are about 2-3 times faster than what I get on our machines
here), even though I thought the normal default for Clp is to use dual simplex.

Cheers,

Kish Shen
Imperial College London
U.K.