Using Iterative Linear Solvers in IDA

[AvivBarnea]

Hello,
I've been using IDA through scikits.odes for quite some time, and thank you for this package. I could never work out how to use non-dense linear solvers, as they fail consistently. What I'm doing really calls for a Krylov solver as the system is stiff, large, but can be preconditioned quite inexpensively.
Could I be guided in employing those solvers? If they are not supported in practice via scikits.odes, I can try to help if guided where and how to see and implement the missing implementation.

The only limitation written right now is the following:

odes/packages/scikits-odes-sundials/src/scikits_odes_sundials/ida.pyx

Lines 879 to 887 in ea608d8

	'linsolver':
	Values: 'dense' (= default), 'lapackdense', 'band',
	'lapackband', 'spgmr', 'spbcg', 'sptfqmr'
	Description:
	Specifies used linear solver.
	Limitations: Linear solvers for dense and band matrices can
	be used only for serial implementation. For parallel
	implementation use lapackdense or lapackband
	respectively.

Thanks again,
Aviv

[c-randall]

I found a couple errors in the Krylov preconditioners for IDA a while back. I submitted #200 to resolve the issues.

There are some accompanying tests in the PR that demonstrate how I was able to get the iterative solvers to work for a simple stiff test problem. Perhaps they could help give you a place to start with your debugging? You'll have to wait for the maintainers to accept the PR and push a new release out to get the fixes though, or create a local branch and build from source with the code from that PR.