[Cbc] cbc executable behaving very differently on different platforms (Linux and Moac OS X)
Stefan Vigerske
stefan at math.hu-berlin.de
Wed Jan 30 12:10:02 EST 2008
Hi,
don't know whether that is a cause of your observations, but maybe you
like to try:
In gcc 3.3.5 there is bug in the optimization where leaf frame pointer
(whatever this is...) are omitted. I normally set the OPT_CXXFLAGS to
"-O3 -fno-omit-frame-pointer -momit-leaf-frame-pointer -DNDEBUG" to get
around it.
Best,
Stefan
Kish Shen schrieb:
> Hi,
>
> I have been having this problem with using CBC (with CLP as the linear
> solver) that was reported by a user of our system. Initially, this was
> from using my own code that interfaces to CBC, but I have now managed to
> reproduce some of the problems with the cbc executable that comes with
> CBC, by feeding it MPS files representing my problems.
>
> I have done this for two problems, both of which are (different)
> formulation of a `stable marriage problem', and both are solved to find
> a feasible solution to the problem (i.e. it has a dummy objective). Both
> problems should be feasible, and indeed when I ran them with CPLEX,
> solutions are given for both. However, when I use the cbc executable (as
> well as my own code interfacing to CBC/CLP), I get different results,
> depending on the platform I ran on.
>
> I have tried this problem on a Linux (x86, 32 bit) box, and also on an
> Intel Mac OS X (also 32 bit). Each machine is able to find a feasible
> solution for one of the problem, i.e. one of the problem is feasible on
> the Intel Mac but is infeasible on the Linux box and vise versa for the
> other problem.
>
> I have noticed from the log of the cbc execution that the problems seem
> to be treated very differently on the two platforms. For example, for
> the same problem, here is what CBC log shows after the processing:
>
> Intel Mac:
>
> mediaLab-MacPro:~/RR20080121 kish$
> ~/Coin-Cbc20071109/i386_macosx/bin/cbc gen.mps
> Coin Cbc and Clp Solver version 2.00.00, build Nov 20 2007
> command line - /Users/kish/Coin-Cbc20071109/i386_macosx/bin/cbc gen.mps
> At line 1 NAME BLANK FREE
> At line 2 ROWS
> At line 1828 COLUMNS
> At line 7733 RHS
> At line 8518 BOUNDS
> At line 11751 ENDATA
> Problem BLANK has 1824 rows, 2848 columns and 9232 elements
> Coin0008I BLANK read with 0 errors
> ...
>
> and after processing:
>
> Cgl0004I processed model has 1578 rows, 2354 columns (2354 integer) and
> 8240 elements
> ...
>
> and a solution is returned:
>
> Result - Finished objective 0 after 4150 nodes and 667378 iterations -
> took 436.20 seconds (total time 436.24)
>
> and for Linux:
>
> kish at bodgerix:~/Bugs/RR20080121> ~/Coin-Cbc20071109/i386_linux/bin/cbc
> gen.mps
> Coin Cbc and Clp Solver version 2.00.00, build Nov 30 2007
> command line - /homes/kish/Coin-Cbc20071109/i386_linux/bin/cbc gen.mps
> At line 1 NAME BLANK FREE
> At line 2 ROWS
> At line 1828 COLUMNS
> At line 7733 RHS
> At line 8518 BOUNDS
> At line 11751 ENDATA
> Problem BLANK has 1824 rows, 2848 columns and 9232 elements
> Coin0008I BLANK read with 0 errors
> ....
>
> but after processing, the model seem to have a lot more rows:
>
> Cgl0004I processed model has 10094 rows, 2329 columns (2329 integer) and
> 25161 elements
>
> and the problem cannot be solved:
>
> Result - Finished objective 1e+50 after 0 nodes and 0 iterations - took
> 48.70 seconds (total time 48.81)
> Total time 48.82
>
> -----
> with very different nodes and iteration for the MIP search...
>
> Any idea if the runs of the same problem on different platform should be
> so different?
>
> The same CBC source is used in both cases (downloaded 9 Nov 2007) from
> the trunk branch, and the same options were given to configure to
> configure the code for compilation. The gcc on the Mac is version 4.0.1,
> and on Linux is 3.3.5. The same MPS file is feed to both.
>
> I can provide the two MPS files for the two problems I have tried.
>
> Thanks in advance for any help/information!
>
> Yours sincerely,
>
> Kish Shen
>
> _______________________________________________
> Cbc mailing list
> Cbc at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/cbc
>
--
Stefan Vigerske
Humboldt University Berlin, Numerical Mathematics
http://www.math.hu-berlin.de/~stefan
More information about the Cbc
mailing list