ESCHER: Expressive scheduling with ephemeral resources

#ephemeral_resources #scheduling_flexibility #scheduling_requirements #Kubernetes #Ray

Meta Info

Presented in SoCC 2022.

Authors: Romil Bhardwaj (UC Berkeley), Alexey Tumanov (Georgia Tech), Stephanie Wang (UC Berkeley), Richard Liaw, Philipp Moritz, Robert Nishihara (Anyscale), Ion Stoica (UC Berkeley).

Understanding the paper

• Goal: support custom scheduling constraints

• Evolvability

  • Monolithic schedulers (Kubernetes, YARN)

    • Applications state resource requirements

    • Scheduler provides a fixed set of supported policies (e.g., affinity)

    • Simple, but hard to evolve

  • Two-level schedulers (Mesos, Omega)

    • Applications implement end-to-end scheduling

    • Highly evolvable, but complex (application must implement a custom scheduler)

• ESCHER

  • Two key abstractions

    • Resource matching scheduler

    • Applications create ephemeral resources and get cluster state at runtime through an API

  • Application -> ESCHER Scheduling Library (ESL) -> Framework Scheduler

  • Add latency (because of RPC call), reduce the implementation burden

  • Implemented in Ray and Kubernetes