Voda Scheduler

is a GPU scheduler for elastic deep learning workloads based on Kubernetes, Kubeflow Training Operator and Horovod.

It offers a system for scheduling elastic training jobs, as well as a number of state-of-the-art scheduling algorithms, that are designed specifically for elastic training to reduce job completion time and increase cluster efficiency. The scheduling algorithms can also be customized at ease.

https://github.com/heyfey/vodascheduler

Features

• Rich Scheduling Algorithms (with Resource Elasticity)
• Topology-Aware Scheduling & Worker Migration
• Microservices architecture
• Node addition/deletion awareness
• Heterogeneous scheduling
• Fault-Tolerance

See the github page for more details.

Architecture Overview

Voda scheduler is designed to be cloud-native and adopts microservices architecture, consisting of several loosely coupled components. It leverages several existing open-sourced projects for reliability, flexibility, and maintainability.