Performance issues in fractional ion balance

[mattngc]

Unfortunately I'm experiencing performance issues with some of the fractional ion balance demos. I've gone back through the various demos and now finding I struggle to run some of the older benchmark cases. The benchmarks I ran in #16 are now taking a very long time.

For example, when I run 'demos/frac_abundance.py' the script now seems to get stuck around iteration 10 lines, 106-108. I think I'll need some help from @CnlPepper to understand why this is happening. @Mateasek reports that these demos are running fine for him.

[mattngc]

If I switch to carbon it works. If I leave it on Neon, it reliably hangs (100% CPU) around iteration 10. I believe there is a numerical issue somewhere.

[Mateasek]

Try to set the max_iter parameter for the lsq_linear which I use to solve the equation here

maybe this is what is iterating endlessly?

or try to step into the solve_ion_balance function and see where this leads? In this function the balance matrix is constructed and solved. It should tell us where it is getting stucked.

[mattngc]

Ok, indeed that does fix the problem. I'm seeing some very peculiar behaviour though. Somehow the execution time is nonlinear with the number of iterations. I'm wondering if that matrix is nearly singular or something like that. I'm finding changing the max_iter parameter from 'max_iter=70' to 'max_iter = 80' is the difference between it running and crashing. Something strange is going on there.

[Mateasek]

quickly searching I found this issue with the scipy function: scipy/scipy#9982

It sounds like what you are experiencing. Do you have the latest scipy? From the issue it seems they have fixed it.

[jacklovell]

I can confirm that the Scipy version seems to be the culprit. Running on freia with Scipy 1.1.0 I get a hang with 100% CPU. Running on my local machine with Scipy 1.3.0 the example runs fine in just a few seconds.