Ray autoscaler is the core engine for running distributed Ray applications and the most used Ray library. Autoscaler runs under the hood with every Ray cluster launched. It manages the entire cluster of nodes based on demand and the user’s application. Autoscaler dynamically determines how many nodes and which types of nodes are running at a given time, how many to launch, how many to terminate, how efficient the application will run and more.
In this talk, Ameer gives a glimpse into Ray autoscaler: the algorithm it uses to schedule resources and the different recent new features added to it including placement groups, directly requesting resources, and multi node types.
Speakers
Ameer Haj Ali