Open joseph-long opened 4 years ago
It seems likely that this is related to https://github.com/numpy/numpy/issues/12753#issuecomment-456903747 but the recommended solution of conda install numpy 'blas=*=mkl'
doesn't seem to change the vectorization of the ufuncs like cos
. (Thinking maybe the Anaconda / Intel patches are the key, I tried conda install -c anaconda numpy
but after a long time waiting for the solve to finish I gave up. This is also what led me to try installing a fresh Anaconda from the latest installer on my account.)
Two users with accounts on the same system reported dramatically different runtimes for a script that made heavy use of NumPy trig ufuncs. I verified this minimal program took approximately 20x longer for user B than user A.
cosprof3.py
A:
B:
Actual Behavior
User B has much slower execution than User A on the same physical computer.
Expected Behavior
User B and User A have approximately equal execution times on the same physical computer.
Steps to Reproduce
The following is the output from running
conda env export --file atrodack_base.yml
from user A's environment, minus references to packages that no longer exist (old OpenSSL, two things related to PyCuLib) that were preventing creating a new environment from the file.atrodack_base.yml
conda env create -f atrodack_base.yml
conda activate atrodack
cosprof3.py
for baseline time valueAnaconda or Miniconda version:
Unknown what versions exactly were used by the users in question, but can reproduce slow behavior with latest miniconda3 and anaconda3 installers on my own account.
Operating System:
conda info
A:
B:
conda list --show-channel-urls
A:
B: