The Issue

I'm running the following numpy benchmark inside of Sage, and comparing it with the results I get by just running Python.

import numpy as np
import time
n = int(10000)
A = np.random.randn(n,n).astype('float64')
B = np.random.randn(n,n).astype('float64')
start_time = time.time()
nrm = np.linalg.norm(np.matmul(A,B))
print(" took {} seconds ".format(time.time() - start_time))
print(" norm = ",nrm)

Output from my Python3 Jupyter notebook:

 took 3.2952768802642822 seconds 
 norm =  999954.1727829538

Output from my SageMath Jupyter notebook:

 took 35.73347020149231 seconds 
 norm =  999976.601519372

Wow. So Numpy runs 10x faster when run inside of a standard Python environment, rather than, e.g. inside of a Sage environment in a Jupyter notebook.

The Cause

I checked the version of numpy installed, especially the BLAS info, and got identical results for both the Sage and standard Python notebook environments. I then checked htop and found the issue: the Python notebook fully utilizes all of the threads on my AMD CPU (all 32 of them) whereas the Sage notebook only uses a single core. This is enough to explain the massive speed difference on my machine.

Is there a way to enable Jupyter's Sage environment to give Numpy access to all of the CPU cores? I tried adjusting the SAGE_NUM_THREADS environment variable to be greater than one but when I launch the Jupyter service and then open a Sage environment SAGE_NUM_THREADS is somehow auto-set back to 1. This bug doesn't show up when I launch the standard Python environment using the same Jupyter service.

Related:

https://ask.sagemath.org/question/445...

Note: Something interesting to note: this slowdown is not observed on my Intel CPU, which uses MKL and runs at the same speed independent of which environment is calling Numpy.