Ilúvatar

Ilúvatar is an open Serverless platform built with the goal of jumpstarting and streamlining FaaS research. It provides a system that is easy and consistent to use, highly modifiable, and directly reports experimental results.

Try it Out

# Dependencies
sudo apt-get install -y curl runc bridge-utils iptables net-tools python3-venv sysstat containerd jq pkg-config cmake gcc g++ libssl-dev pkg-config libprotobuf-dev
# Rust, more at https://www.rust-lang.org/tools/install
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

# Install non-apt dependencies and build Ilúvatar, assumes python3 exists
cd src/Ilúvatar/docs/examples
./sample_setup.sh
sudo ../../target/debug/iluvatar_worker -c ../../iluvatar_worker/src/worker.dev.json &
worker_pid=$(echo $!)

Register a function with the worker.

../../target/debug/iluvatar_worker_cli --address "127.0.0.1" --port 8079 register --name "hello" --version 1 --image "docker.io/alfuerst/hello-iluvatar-action:latest" --memory 128 --cpu 1

Invoke the newly registered function, passing custom arguments.

../../target/debug/iluvatar_worker_cli --address "127.0.0.1" --port 8079 invoke --name "hello" --version 1 -a name=`whoami`

Kill the worker running in the background.

sudo kill --signal SIGINT $worker_pid

You can run more advanced scenarios here.

Why use Ilúvatar for research?

Ilúvatar provides the features and guarantees we find to be necessary to conduct high-quality, reproducible, and open research.

1. Load generation a first-class citizen of the system, making results easier to capture and analyze.
2. Fast, predictable, jitter-resistant function execution for reliable and repeatable experimentation.
3. An architecture designed to allow testing on a single worker node, or a cluster of load-balanced nodes.
4. The ability to run a load scenario on a live system or as a simulation, using the same codebase.
5. An extendible FaaS platform built for an open-source research community.

Details about the implementation can be found in the paper below. How to use the various features and jump-starting your research can be found in the Ilúvatar documentation.

A popular open-source platform used in research is OpenWhisk, but we have found it has high overheads under notable load, caused by a variety of factors. The scalability of OpenWhisk vs our Ilúvatar can be seen here.

We can ensure a constant 1-3 ms overhead on invocations at significant load, whereas OpenWhisk sees high and variable overheads, on the same hardware. It also does not have the research-first features implemented in Ilúvatar.

See Ilúvatar in action for research right here

More documentation / Questions

Ilúvatar supports a large variety of customization in configuration and setup, and methods of load generation and experimentation. It is easy to create custom functions that run new Python programs, information is in this document. Detailed documentation can be found closer to the code.

If you have questions, want help working with Ilúvatar, or anything else, we have a Discord server. Definitely gives faster answers than opening an issue for a question.

Why 'Ilúvatar'?

The name 'Ilúvatar' comes from J.R.R. Tolkien's mythology, published in The Silmarillion, a history of the world prior to the events of his Hobbit and Lord of the Rings books. Ilúvatar is the creator of the world, and orchestrated its form and development. We don't see our platform as being the ultimate, final, or conclusive FaaS platform. But as a FaaS platform, it controls and directs where and how Serverless functions are created and executed, across a possibly great variety of heterogeneity and network distance. Thus the inspiration for its name.

Citation

If you use, extend, compare against, etc., Ilúvatar, please reference our HPDC 2023 paper in your work.

Alexander Fuerst, Abdul Rehman, and Prateek Sharma. 2023. Ilúvatar: A Fast Control Plane for Serverless Computing. In Proceedings of the 32nd International Symposium on High-Performance Parallel and Distributed Com- puting (HPDC ’23), June 16–23, 2023, Orlando, FL, USA. ACM, New York, NY, USA, 14 pages. https://doi.org/10.1145/3588195.3592995

Acknowledgements

This work has been funded by grants from the National Science Foundation, including the ICICLE AI Institute (OAC 2112606)