[Cbc] memory leaks in time-constrained parallel cbc-based solver

[John Forrest]

Pietro,

Will look into it after INFORMS.

John

On 08/11/14 11:56, Pietro Belotti wrote:
> Hello. I am using libCbc to build a convex MINLP solver for a specific
> application. The solver is based on Cbc's branch-and-bound and a cut
> separator. I have an incomplete problem at every node, since the
> linear cuts are a relaxation of a nonlinear (convex) constraint. This
> is done through
>
>    OsiBabSolver solverChar;
>    solverChar. setSolverType (3);
>    dsbb. solver () -> setAuxiliaryInfo (&solverChar);
>
> When configuring with --enable-cbc-parallel, I can run this program in
> parallel, but two problems occur:
>
> 1) if a time or node limit is given, there are memory leaks at the end
> (see the two attachments valg-out2-200nodes.txt and
> valg-out2-10s.txt), but no leak happens when reaching optimalitiy (see
> valg-out2.txt); most of these seem to be cuts added through addCuts()
> at CbcNodeInfo.cpp:354. I have removed non-Cbc output from these
> files.
>
> 2) The parallel version seems to use more memory than the single-CPU
> version at a given time. A back-to-back solve shows that the parallel
> version uses 2x the memory of the sequential one even when they are
> exploring the same number of nodes; memory occupation is the same
> until the BB starts.
>
> It should be emphasized that this solver generates a large number of
> cuts, and for some instances after 500,000 nodes it is killed after
> using the 240GB memory limit. I could not replicate this with the
> console version of parallel cbc with time/node limit on a MILP, so I
> suspect I might be doing something wrong. Has anything like this
> happened in the past?
>
> Thanks,
> Pietro
>
>
> _______________________________________________
> Cbc mailing list
> Cbc at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/cbc

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/cbc/attachments/20141112/e8c4b7f6/attachment.html>