cedana-cli

README

cedana-cli

Cedana is a framework for the democritization and (eventually) commodification of compute. We achieve this by leveraging checkpoint/restore to seamlessly migrate work across machines, clouds and beyond.

This repo contains a self serve CLI tool to allow developers to experiment with our system. With it, you can:

• Launch instances anywhere, with guaranteed price and capacity optimization. We look across your configured providers (AWS, Paperspace, etc.) to select the optimal instance defined in a provided job spec. This abstracts away cloud infra burdens.
• Leverage our client code (installed on every launched instance) to checkpoint/restore across instances, unlocking reliability, increased performance gains and decreased price.
• Deploy and manage any kind of job, whether a pyTorch training job, a webservice or a multibody physics simulation.

To access our managed service, contact founders@cedana.ai

Usage

Cedana consists of the client code (found here) running on compute in the cloud (or anywhere else) and the orchestration/daemon, which runs on your local machine.

To build from source: go build

To run: ./cedana-cli

If you prefer to install from a package manager, we push to packagecloud and have a homebrew tap. Check out the documentation for instructions.

Documentation

You can view the official documentation here.

Demo

https://www.youtube.com/watch?v=KC4STzSQ_DU

(Note: The video is sped up for brevity to show how a CPU-bound PyTorch training job can be migrated between instances in realtime).

Todos

We're working on building out a public roadmap. Until then, here's a few of the highest priority todos:

• Add more cloud providers to arbitrage between
• runc container checkpointing
• Advanced optimizaiton strategies to pick and migrate work between clouds
• Way more tests
• GPU checkpointing
• Simulation environment for rapid checkpoint/migrate
• Kubernetes and cluster formation support
• Batch compute paradigms
• SLURM feature parity

For checkpoint/restore specific work, refer to the README in the client code repo.

Contributing

See CONTRIBUTING.md for guidelines.