AWS spot GPU instances are a pretty good option for the fast.ai classes.Their p2.xlarge spot instance (K80 GPU, 4 CPU, 60GB RAM) is usually priced at only $0.3/hr, 3x cheaper than the normal p2.xlarge instance. We tested it out but seems like there’re a few problems with hosting jupyter notebook on spot instances
- configuring ssh forwarding for jupyter notebook is an overhead for some classmates.
- the spot instances get shut down from time to time and we could lose our code & whole environment setup.
Hosting jupyter notebook on AWS spot instances is supposed to be easy for everyone! We built a few things around AWS optimized for fast.ai course materials:
- a low-cost persistent storage connected to your spot instance so even if it gets stopped, your code, data and course setup is still there.
- a task scheduling system to re-schedule the jupyter notebook on spot instance automatically when there’s spot interruption.
You can check out more details here in this blog http://blog.snark.ai/posts/2019-02-24-jupyter_on_spot.html. Let me know how we can improve the service.