A Full-Scenario Colocation of Workloads Based on Kubernetes

Explore a comprehensive conference talk on maximizing resource utilization through workload colocation based on Kubernetes. Learn how to effectively combine online services and offline jobs to improve efficiency and reduce costs. Discover techniques for resource prediction, isolation, interference detection, and offline eviction that enable optimal resource usage without compromising online service SLOs. Gain insights into using eBPF for kernel-level metric collection to detect interference when latency metrics are unavailable. Examine the implementation of these techniques on native Kubernetes, supporting various scenarios including containerized and non-containerized services, as well as Kubernetes and Hadoop ecosystem jobs. Understand the real-world impact of this approach, as demonstrated by Tencent's deployment across 40,000+ machines, resulting in a 15% average increase in utilization and significant cost savings.

Intro
Why colocation
Make the colocation better
Colocation on K8s — Caelus
Principles on kubernetes
Full-scenario colocation
Resource prediction
Prediction algorithms
Cgroup hierarchy
Resource isolation
Interference detection
Resource load detection
RT detection
Function detection
Interference handling
Improve resource utilization
Run more offline jobs
Results