hal2069 / Faasm

High-performance stateful serverless runtime based on WebAssembly

Faasm

Faasm is a high-performance stateful serverless runtime.

Faasm provides multi-tenant isolation, yet allows functions to share regions of memory. These shared memory regions give low-latency concurrent access to data, and are synchronised globally to support large-scale parallelism.

Faasm combines software fault isolation from WebAssembly with standard Linux tools, to provide security and resource isolation at low cost. Faasm runs functions side-by-side as threads of a single runtime process, with low overheads and fast boot times. The underlying WebAssembly execution and code generation is handled by WAVM.

Faasm defines a custom host interface which extends WASI to include function inputs and outputs, chaining functions, managing state, accessing the distributed filesystem, dynamic linking, pthreads, OpenMP and MPI.

A preprint of our paper on Faasm can be found here.

Quick start

You can start a Faasm cluster locally using the docker-compose.yml file in the root of the project:

docker-compose up --scale worker=2

Then run the Faasm CLI, from which you can build, deploy and invoke functions:

# Start the CLI
./bin/cli.sh

# Upload the demo "hello" function
inv upload demo hello

# Invoke the function
inv invoke demo hello

Note that the first time you run the local set-up it will generate some machine code specific to your host. This is stored in the machine-code directory in the root of the project and reused on subsequent runs.

More information

More detail on some key features and implementations can be found below:

• Usage and set-up - using the CLI and other features.
• C/C++ functions - writing and deploying Faasm functions in C/C++.
• Python functions - isolating and executing functions in Python.
• Rust functions - links and resources for writing Faasm Rust functions.
• Distributed state - sharing state between functions.
• Faasm host interface - the serverless-specific interface between functions and the underlying host.
• Kubernetes and Knative integration- deploying Faasm as part of a full serverless platform.
• Bare metal/ VM deployment - deploying Faasm on bare metal or VMs as a stand-alone system.
• Tensorflow Lite - performing inference in Faasm functions with TF Lite.
• API - invoking and managing functions and state through Faasm's HTTP API.
• MPI and OpenMP - executing existing MPI and OpenMP applications in Faasm.
• Local development - developing and modifying Faasm.
• Faasm.js - executing Faasm functions in the browser and on the server.
• Threading - executing multi-threaded applications.
• Proto-Faaslets - snapshot-and-restore to reduce cold starts.

Experiments and benchmarks

Faasm experiments and benchmarks live in the Faasm experiments repo.