Server Partitioning Without VMs - for Flexibility and Performance
summary
Shows a hands-on method to partition a 128‑core, two‑NUMA Intel server for containers using NRI plugins and the balloons policy inside containerd (no VMs). Demonstrates namespace‑based balloons that map pods to exact CPU sets, inflating dynamically and keeping workloads on separate L2 caches (example pods used cores 17→81 and 19→83), enabling cache‑aware isolation, predictable latency, and full‑node fallback when balloons deflate.
AI-written summary · about
chapters (7)
- 0:00Server partitioning intro
- 3:10Kubernetes node limits
- 10:45Node Resource Interface
- 13:33Cache/NUMA partitioning
- 16:44Live cache isolation demo
- 29:21Recap and plugins
- 30:50Q&A
related talks
41mAn Essential Relationship between Real-time and Resource Partitioning
33mLinux Container Internals
20mGet More Computing Power by Helping the OS Scheduler
38mOn the Way to Safe Containers
25mYou Can’t Handle the Root! Container Isolation...
43mLinux Container Performance Analysis
49mThings I Wish I Had Learned Earlier About Containers: Lessons from a Seasoned Linux...
31mHigh Density VMs in OpenShift: The Journey to Native K8s Swap