Skip to content
/ melody Public

Online scheduling framework for edge ML inference on Kubernetes

Notifications You must be signed in to change notification settings

wanziyu/melody

Repository files navigation

Melody: ML Edge Inference with Dynamic Scheduling

Melody is a framework for intelligent scheduling for edge computing ML inference instances on Kubernetes.

Overview

Melody uses Deep Reinforcement Learning as the scheduling algorithm, use prometheus to monitor Edge computing resources and input metrics to DQN, and obtain the optimal scheduling strategy.

Features

key benefits include:

  • Support auto-scheduling ML inference instances between edge nodes.
  • Continuously balance the computing resources (cpu, memory) on edge nodes.
  • Equipped with reinforcement learning algorithm to obtain the balanced scheduling strategy.

Custome Resource Definition(CRD)

  • Inference define the ML inference jobs, it observes the scheduling decesion CRD, and dynamic adjust the resource limit and request.
  • Scheduling defines the scheduling strategy by communicating with RL algorithm server, it defines the optimal edge node for each ML serving instance Job.

Get Started

About

Online scheduling framework for edge ML inference on Kubernetes

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published