[Coin-lpsolver] how large scale problems clp can solve?

[Michael Hennebry]

On Sun, 27 Jan 2008, Jinwei Gu wrote:

> I am new to CLP. I am wondering what is the limit of the scale of the
> problems that CLP can solve.
>
> My research problem has about 1 million unknowns (i.e. columns), and 4
> million constraints (i.e. rows). The matrix is very sparse. Each of the rows
> only has about 100~400 non-zero values. I am wondering whether CLP can be
> used to solve this problem?
>
> I ran several test problems and CLP works much faster than glpk which I used
> before. Actually CLP is the only open-source library I found that might be
> able to solve that large scale problems. Is it just limited by my machine
> memory? Is there a difference in terms of the scale limit, if I use simplex
> method (primal/dual) or the barrier method to solve?

Both GLPK an CLP are limited by the amount of memory
available for data structures in a single process.
You need at least 4 million * 100 * 8 > 3 billion
bytes just to store the constraint matrix.
Other data structure take up more than twice that.
You need a computer and OS that will let you address more than 4GB.
A 32-bit machine/OS won't work.

To the best of my knowledge, both CLP and GLPK
will both compile and run on 64-bit machine/OSs.
How much does 16GB of ram cost these days?

Of course, if the basis factorization turns out to be dense,
you are into the realm of the terabyte.

-- 
Michael   hennebry at web.cs.ndsu.NoDak.edu
"Those parts of the system that you can hit with a hammer (not advised)
are called Hardware;  those program instructions that you can only
curse at are called Software."