Update documentation for setting Python environment #778
Conversation
| @@ -77,7 +77,7 @@ This page goes through the Python environment setup steps in more detail and wit | |||
| WARNING: The Python readline extension was not compiled. Missing the GNU readline lib? | |||
| WARNING: The Python sqlite3 extension was not compiled. Missing the SQLite3 lib? | |||
|
|
|||
| 2. Install the required version of Python via `pyenv`, and _remember to enable it as a shared library_ so Theano can call into it: | |||
| 1. Install the required version of Python via `pyenv`, and _remember to enable it as a shared library_ so Theano can call into it: | |||
There was a problem hiding this comment.
Is the "shared library" part also needed on macOS? I forgot! I suspect that I installed 2.7.16 locally without it.
|
|
||
| ```bash | ||
| pip install --upgrade pip setuptools virtualenv virtualenvwrapper virtualenv-clone wheel | ||
| ``` | ||
|
|
||
| 3. Install OpenBLAS 0.3.5 or later. | ||
| 1. Install OpenBLAS 0.3.5 or later. |
There was a problem hiding this comment.
Maybe add that openblas 0.3.7 (as installed by brew install openblas) works fine on macOS but not inside Docker on macOS unless you compile it with NO_AVX2=1. Maybe this is too complicated to even get into. OpenMathLib/OpenBLAS#2244
| 1. Add the following line to your bash profile. | ||
| ``` | ||
| export OPENBLAS_NUM_THREADS=1 | ||
| ``` |
There was a problem hiding this comment.
This is a good thing. (And switching to plain 1. everywhere is also good.) But do we always do this? I do not know. Or always configure it this way and sometimes override it with a different local value?
FYI, Intel's alternative, Math Kernel Library (MKL) implementation has a Thread Building Blocks (TBB) feature to avoid oversubscription of threads while not limiting itself to one thread per process. The last time I tried it, it was slower than Openblas. It turns out there's a new release for 2020 and there were several releases in 2019, so maybe we should retest it. See #36
There was a problem hiding this comment.
I've been using the export for awhile. Maybe it's linux specific but you get much better performance. Maybe we could say it's optional or add some info about why you want to add it.
Testing MKL again could be worthwhile but do you think they're optimized for intel processor instruction sets/architecture and not so much for AMD/others?
There was a problem hiding this comment.
Yuk. Very good point that we'd have to test on at least one Intel CPU and at least one AMD CPU, maybe more, and the Docker-on-Mac-on-Intel case. At least it only takes replacing some pips -- on each of those system installations. I'm not excited to test it again.
| 1. Add the following line to your bash profile. | ||
| ``` | ||
| export OPENBLAS_NUM_THREADS=1 | ||
| ``` |
There was a problem hiding this comment.
I've been using the export for awhile. Maybe it's linux specific but you get much better performance. Maybe we could say it's optional or add some info about why you want to add it.
Testing MKL again could be worthwhile but do you think they're optimized for intel processor instruction sets/architecture and not so much for AMD/others?
This adds a step in the documentation that sets $OPENBLAS_NUM_THREADS=1, which @tahorst pointed out as a solution for slow simulation times in #448. Adding this cut down the simulation time on my local machine from 16 minutes to 10 minutes even when running a single simulation.