Improve MNIST and benchmark

[laszewsk]

• [x] make sure code uses GPUs
• [ ] add probe or flag to make it clear that we use GPU or not
• [ ] include options to prg with click as easuer in this case

Benchmark

• [ ] Thomas
• [x] Robert
• [ ] Jake

[rknuu]

Benchmarks from my runs are below; all from the same system.

When I ran the code, it was clear that tensorflow and keras were looking for my graphics card (there were log messages such as 2022-02-13 02:22:44.868704: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1525] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 13584 MB memory: -> device: 0, name: NVIDIA GeForce RTX 3080 Laptop GPU, pci bus id: 0000:01:00.0, compute capability: 8.6

WSL2 Ubuntu (3.8.8) Note - execution illustrates issues where the kernel was built without NUMA support.

• Real: 44.560s
• User: 26.837s
• Sys: 13.598s

Windows CPython (3.9.7, with pip package)

• Real: 20.490s
• User: 0.015s
• Sys: 0.000s

Windows Anaconda (3.9.7, with pip packages)

• Real: 20.889s
• User: 0.000s
• Sys: 0.015s

Windows Anaconda (3.9.7, using conda-native packages) Failed at first (missing keras package; installing keras fixes this)

• Real: 12.384s
• User: 0.000s
• Sys: 0.015s

Windows Anaconda (3.9.7, with conda-native packages and using tensorflow-gpu package)

• Real: 21.667s
• User: 0.015s
• Sys: 0.000s

[laszewsk]

can you add

a) python.org 3.10.2 b) python.org 3.10.2 compiled from source with optimizations c) python-conda 1.10.2 d) can you make sure it runs on GPU and CPU e) can we make a script so that other can just run the script. (e.g. bash) we run it via gitbash

I suggest to do package download to ~/Downloads and environments such as ~/BENCH-x.yzz/

I like to add this framework to cloudmesh-benchmark (which I will create.