Single file implementation of a hardware monitor to be used with tensorboard, wandb, and other logging solutions that want to work with a dictionary. Put it in your codebase and use it however you like.
A very basic hardware monitor based on psutil that enables tracking of:
- network bandwidth
- disk read/write bandwidth
- disk read/write counters
- context switches
- average CPU load (1,5,15 min)
- memory utilization
- process memory (resident, virtual, lib, etc.)
If a GPU is available, over pynvml:
- framebuffer memory
- bar1 memory
- gpu/memory "utilization"
- temperature
- power
- throttle reasons
If a GPU is available, and torch:
- all of the stats in
torch.cuda.memory_stats() - average (small, large, all) segment size
- average block size
- average inactive block size
- allocation rounding overhead
- memory fragmentation
- An example application with torch
- Write some tests
- Better decapsulation between
psutil,pynvmlandtorch
If you found this useful and want to cite the work, please use the following bibtex:
@software{Isenko_Basic_Hardware_Monitor_2023,
author = {Isenko, Alexander},
license = {MIT},
month = sep,
title = {{Basic Hardware Monitor}},
url = {https://github.com/cirquit/py-hardware-monitor},
version = {0.1},
year = {2023}
}