Install tensorflow with GPU support on Ubuntu 20.04 and anaconda perfomance observations

[Gregor von Laszewski]

All:

TENSORFLOW ON UBUNTU 20.04

=====================

As you may know tensorflow does not come for Ubuntu20.04 and is at this
time not yet supported.

Previously, I distributed instructions of some 30+ steps that included
the compilation of tensorflow by hand on Ubuntu which was super
complicated including a custom gcc+ compiler install. I have removed
these instructions and replaced them.

The reason is that the new tensorflow release  has the compile isse
resolved. I am documenting how to  still use cuda 7.6.5 drivers which is
used by tensorflow and  is super useful for us as Romeo uses the same
driver version. See:

https://github.com/cloudmesh-community/book/blob/master/chapters/ai/tensorflow-on-ubuntu-2004.md

ANACONDA PERFORMANCE ISSUES ON ROMEO

=============================

Also I found out that using anaconda is ***questionable*** on Rome, as
if you use a regular python  (compiled with optimizations) it is much 
faster (at least for me). For example when using anaconda  on Romeo my
program was running for 2 hours while running it with my optimized
python it only took 20 minutes. So before you all run anaconda, I
suggest you explore performance improvements you gain when using proper
python. This is a meaningful difference as you are not only effecting
your own wait, but also others that may want to use the GPUs or CPUs on
Romeo.


PYTHON 3.8.5 ON ROMEO

================

Allan has also installed 3.8.5 but I had no time to benchmark this, so
if you have time, lt us know what you find. Due to python 3 venvs are
trivial to create with

python -m venv ....



If you have other tips, please let me know, or create a pull request
with your improvement suggestions

Gregor

PS: If you tried the new drivers with tensorflow, let me know

[Allan Streib]

A couple of comments:

The version of Anaconda on Romeo is at this point relatively
old. Perhaps worth an update?

Python 3.7 and 3.8 versions on Romeo were not compiled with any
optimizations. It was a very standard configure; make; make install.

Allan

--
Allan Streib
Indiana University
Luddy School of Informatics, Computing, and Engineering
Digital Science Center | Intelligent Systems Engineering | FutureSystems