Massive performance regression at #1b7f414

[jarvist]

Massive performance regression at #1b7f414

[Neutrino155]

Hmm, that's not good. I'll have a look.

[Neutrino155]

Oh, I think that the issue is that the Optim minimisation is showing the full trace.

[Neutrino155]

The method is also different as I removed minimisation constraints (Fminbox etc.) whilst generalising to multiple variational parameters. This is no longer required as I found a way to add the constraints back.

[Neutrino155]

So I've fixed some parts of this, but I still see a massive slowdown for the Julia v1 tests. Julia 1.7 runs within a couple of minutes. Must be one of the dependencies.

[jarvist]

Interesting that all the slow down is in 'MultiplePhonons' code, do you have an idea of which part of the code is now running so slow? If we can avoid using that temperamental branch, it may be for the best.

[Neutrino155]

The performance regression has been countered. I think the issue, in the end, was a combination of things:

1. Multithreading combined with the variation code that minimises the multiple phonon version of the polaron energy. I am not sure why this combination grinds to a halt. Potential data race happening?
2. I previously removed rtol in the quadgk integrations in the memory function code. This means that the integrals take significantly longer to compute. I think it is okay to include this convergence check at this level of the code.

A note on the memory function calculation. This is used for the complex impedance and mobility calculations. I tried different relative tolerances for the memory function integral and saw little to no difference in the impedance and mobility for rtol=1e-4 compared to lower tolerances. However, limiting the tolerance gives a considerable speed boost. rtol=1e-3 is okay, but sometimes numerical instabilities appear in the mobility at this tolerance so I settled on rtol=1e-4.

Edit: Just another note that potential data race in (1) seems to happen in Julia version 1.0 and 1.8, but not Julia 1.7.