Re: gnumeric's solver and parallel computing

[Frédéric Parrenin <parrenin ujf-grenoble fr>]

The current solver works with 600 variables (it takes about two hours) but I am planning to solve 5-10 times more and I am not sure it will be feasible.

Frédéric Parrenin

2013/2/4 Frédéric Parrenin <parrenin ujf-grenoble fr>

I am playing with about 600 variables.

Frédéric Parrenin

2013/2/4 Morten Welinder <mortenw gnome org>

> I also noticed that the newton iterations converge a lot more quickly than
> the Rosenbrock iterations, at least in some circumstances.
> Would it be possible to make an option for the algorithm which is chosen?

That's not really feasible interface-wise: the current solver bounces
back and forth between the Rosenbrock steps and Newton steps.

Note that when Newton iteration is attempted basically for as long
as it improves the situation.  In other words, there would be no way
of continuing if you were just using just Newton iteration.

How many variables are you playing with?  We're recomputing the
NxN Hessian for every Newton step.  that might be expensive with
lots of variables.

M.