[Coin-ipopt] Re: ipopt

[Andreas Waechter]

Hi Jason

> I am trying to use your ipopt for my project, which, for a moderate problem,
> can easily have 20 million variables. Do you think IPOPT can handle so large
> system, and do you have an estimation of the cpu and memory it requires on a
> linux of 1.5G CPU?

I don't know if Ipopt could handle a problem of that size - it is certain
one order of magnitute larger than what I have tried so far (but of
course it would be great to try it... :).

Of course, all this to a large degree also depends on the structure of
your problem and what fill-in it introduces.  It is impossible to say
anything more than just wild guessing without knowing more about the
problem (constraints?  How dense is the constraint Jacobian, and how dense
is the Hessian of the Lagrangian?  What is their structure?  How many
degrees of freedom?  Are second derivatives available?  How are the
problem functions computed? [AMPL would certainly NOT work])

The bottleneck is the solution of the linear system for computing the
search direction in Ipopt.  Currently, the default interfaces are for
Harwell routines (MA27 and MA57), and I have no idea if they could handle
problems of that size.  I don't even know if there are other direct linear
solvers (for solving indefinite systems and that report the number of
negative/positive eigenvalues of the factorized matrix) that can solve
large matrices of that size.  In the new version of the code we want to
offer interfaces to a few linear solvers, also some with parallel
computing capabilities...

In any case, you would most probably need a 64bit system anyway, since the
memory consumption would probably a quite large - and it would take quite
a while on a single CPU machine.

Maybe a reformulation of your problem can reduce the number of
variables...

A vague answer to a vague question... :)

Cheers

Andreas