[Clp] [Coin-lpsolver] Best way to bring an instance into CLP?

[Lasse Kliemann]

* Message by -John J Forrest- from Fri 2008-01-18:
 
> If memory is an issue the most efficient way takes more effort but is 
> doable if you have a reasonable (over)estimate of numbers of rows, columns 
> and elements.  I have updated Clp/examples/addRows.cpp to create a model 
> from scratch.  The example matrix is built by column.  You could build by 
> row and then transpose before handing to model.
> 
> Hope the example is clear enough.

Hello John, finally I found time to go back to that problem. I implemented it 
the way you demonstrated in the example. It became perfectly clear from the 
example how to do it. Thanks!

Unfortunately, my problem persists. It appears that I did not look closely 
enough before to notice that building the instance is in fact not the 
problem, but solving it. No matter which of the two build methods I use 
(addColumn or the manual way), at a certain threshold of instance size
a call to model.initialSolve() gives an out-of-memory error. Judging from 
runs on smaller instances, there seems to be a memory peak directly after 
the call to initialSolve(). This seems to be due to presolve, for it goes 
away when I do it like this:

   ClpSolve options;
   options.setPresolveType(ClpSolve::presolveOff);
   model->initialSolve(options);

In fact, I can solve slightly larger instances without presolve.

Without presolve, for instances larger than that slightly higher threshold, I 
do not get out-of-memory errors, but:

   Clp3002W Empty problem - 3000 rows, 0 columns and 0 elements

The solver continues, but does not terminate (in the expected time).

Now, in the example, you use model.dual(). When I try that, the solving 
process is much slower, about a factor 30 to 40, but it seems to continue on 
larger instances without warning messages or error.

I also tried

   ClpSolve options;
   options.setDoDual(true);
   model->initialSolve(options);

but that doesn't seems to make a difference compared to just calling 
initialSolve() without options.

I'm using https://projects.coin-or.org/svn/Osi/stable/0.98.

There is one more thing: in cases where the program aborts with an 
out-of-memory error, only about 40% of total memory is used. The system is 
Linux 2.6.24.x, and there is no ulimit:

$ ulimit -a                                                                                                                   
coredump(blocks)     0
data(KiB)            unlimited
file(blocks)         unlimited
flocks               unlimited
lockedmem(KiB)       32
memory(KiB)          unlimited
nofiles(descriptors) 1024
processes            32768
stack(KiB)           unlimited
time(cpu-seconds)    unlimited
vmemory(KiB)         unlimited


Thank you a lot for advice on any of these issues!
Lasse

-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 835 bytes
Desc: not available
URL: <http://list.coin-or.org/pipermail/clp/attachments/20080311/a38a5f0c/attachment.sig>