YATSM >100% CPU utilization with conda

[ceholden]

Looks like the NumPy included in conda can do multithreaded computation for linear algebra routines by linking to OpenBLAS. You can disable this by setting OPENBLAS_NUM_THREADS=1 as an environment variable, but having all my jobs killed by our cluster's process reaper gave me quite the fright before I found the solution to turning it off!

If we linked against the Intel MKL, we'd also have MKL_NUM_THREADS to set.

Proposed solutions:

1. Document it and hope people find it
2. By default, have yatsm set the environment variables for multiprocessing to just use one thread. Use some --num_threads optional argument in the YATSM cli to turn multiprocessing on.

I'm actually in favor of option 2, but it seems pretty oppressive. Any thoughts @valpasq or @bullocke ?

[parevalo]

I think option two makes a lot more sense. Better to go with the safe option and give an option to override it if necessary, no?

[ceholden]

Ignore the previous reference to a commit. I had the wrong issue number

[ceholden]

Commit 6bf7371 doesn't solve the problem. The idea is right, but it looks like we need to set the environment variables BEFORE numpy is ever imported.

This means that we need to import numpy even after we load all of the CLI program entry points (e.g., map, line, train, etc.). This could be a whole lot of work that makes things very messy...