[Ipopt] Memory / Storage requirements when solving with "exact" vs "limited-memory" hessian approximations

[Mehmet Ersin YUMER]

Hi,

I am looking for a way to roughly estimate the required memory size when
using an in-core linear solver (ma27, ma57, etc.) with ipopt, and required
disk storage size when using an out-of core linear solver (ma77).

If:
number of variables = n
number of equality constraints = m
number of inequality constraints = k
number of nonzero in constraint jacobian = j
number of nonzero in lagrangian hessian = h

Is it correct that when using the "exact" hessian, the dominant memory
requirement arises from the search direction matrix, and it is a dense
(n+m)*(n+m) matrix? Or is it more involved than this?

How many of these matrices are stored during the execution?

Is there any difference in size when using an out-of-core linear solver
other than the fact that the matrix is now stored on disk?

What about the "limited-memory case"?


Thanks in advance for your help.

PS: For my problem the values are roughly as follows:
n = 1M
m = 50K
k = 0
j = 1M
h = 500K


-- 
Mehmet Ersin Yumer
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/ipopt/attachments/20130124/bd2775a5/attachment.html>