Ray: A Distributed Framework for Emerging AI Applications

submited by

Style Pass

2021-07-04 16:30:04

This week I decided to revisit a paper from the 2018 edition of OSDI (Operating Systems Design and Impelementation). In the coming few weeks, I will likely be reading the new crop of papers from HotOS 2021. If there are any that look particuarly exciting to you, feel free to ping me on Twitter!

Ray is a thriving open-source project focused on “providing a universal API for distributed computing” - in other words, trying to build primitives that allow applications to easily run and scale (even across multi-cloud environments), using an actor-like framework. There are a few exciting demos which show how easy it is to parallelize computationThe demos are from Anyscale, a company founded by several original authors of the paper. Apache Spark is to Databricks, as Anyscale is to Ray. . The idea that (somewhat) unlimited cloud resources could be used to drastically speed up developer workflows is an exciting area of research - for a specific use case see From Laptop to Lambda: Outsourcing Everyday Jobs to Thousands of Transient Functional Containers.

While the current vision of the project has changed from the published paper (which came out of Berkeley’s RISELabThe RISELab is the “successor to the AMPLab”, where Apache Spark, Apache Mesos, and other “big data” technologies were originally developed) ), it is still interesting to reflect on the original architecture and motivation.