On Mon, Nov 17, 2008 at 2:39 PM, Jonathan Guyer <[EMAIL PROTECTED]> wrote: > >> In my many attempts to sort this out, I varied the solver I was using >> and noted this simulation solves when I select the 'linearPCG' solver, >> but never converges when I use 'linearGMRES', and returns NaNs when I >> use the 'linearCGS' solver. I was surprised that the solver choice was >> so important for such a simple problem. Would you have expected this >> behavior? Does this indicate somethings wrong with pysparse, at least >> on my computer? What do I need to do to try the scipy sparse solvers >> to see how they perform? > > I'll let Daniel field your solver questions. I see the same results you do, > but note that PCG is the default.
I think that the CGS solver in pysparse is broken. The trilinos CGS solver seems to work fine. I've known for some time that the pysparse CGS solver was not right, but just haven't got round to fixing it. I think we should. In the meantime, - we should deny it as an option in fipy for now - add a test in pysparse to show the error - try and fix it I am not sure about GMRES. GMRES is more dependent on parameter choices, again we need to look deeper at this. I recommend benchmarking against the LU solver for all your problems to be certain that things are working correctly. You might want to try and install trilinos as that gives a wider and better choice of solvers. I use the trilinos solvers frequently. I can give you more details if you are interested. As for interfacing with the scipy solvers, we haven't dealt with that yet. I don't believe it is a particularly difficult task considering that all the structure is already there. If you are felling brave then go for it. Sorry I can't be more helpful. -- Daniel Wheeler
