Experiment with PARDISO solvers from MKL

[ViralBShah]

Now in BB as well:

JuliaSparse/Pardiso.jl#56

[ViralBShah]

@ranjanan Can you note your findings here?

[ViralBShah]

@ranjanan reports:
On Antarctic, MKL Pardiso about 20% faster than CHOLMOD using the defaults. Antarctic has 20 cores, which PARDISO is able to utilize.

[ViralBShah]

[ViralBShah]

@vlandau Can you try out using the MKL Pardiso solver in OmniScape? Does it help?

There's a pardiso branch that perhaps may be easy to try. #225

[vlandau]

Definitely! So this is a different method from CHOLMOD, or is it cg+amg using this Paradiso method. Also, it works for advanced mode presumably?

[ViralBShah]

Yes, it should work everywhere. It is just an alternative solver - but it is a sparse direct solver like cholmod. Probably need to set number of threads to 1 in parallel mode (by looking at Intel MKL docs).

[vlandau]

Does that mean I can't run multiple solvers in parallel with this method in Omniscape? (I use multithreading)

[ViralBShah]

You should be able to. Generally Pardiso will try to use all the cores, so you have the over subscription problem. So some environment variable probably should be set so that each Julia process runs pardiso only on one core, when in parallel.

[vlandau]

Okay I see -- sort of similar to over subscription due to BLAS parallelization. I'll look into this!