Honey, I shrunk the pods (and everything else)
summary
Kata Containers runs each pod as a VM and currently adds a fixed memory overhead of about 350–650 MB per VM pod, which grows with I/O; booting a new VM takes roughly ~7 s versus ~0.3 s for runc, though multiple containers in one pod amortize cost. Replacing the Go shim with shim v2 cuts per-pod memory dramatically. Storage moved from 9p to virtiofs with optional DAX sharing (better throughput, shared page cache), while SR-IOV/DPDK is the path for near-native networking.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:06Virtualization overhead
- 3:11Memory density in cata
- 6:23Disk cache and DAX sharing
- 9:08Memory overhead in cata 2.0
- 9:51Boot time comparisons
- 11:15Disk I/O performance
- 13:43Networking overhead
- 15:49SR-IOV optimizations
- 16:10Resource trade-offs
- 17:38Q&A
related talks
25mOne Size Fits None: Lightweight Workload-Specialized Confidential Containers
39mMore Than Secure: Containerd + KataContainers as Kubernetes Runtime
34mPod Through a Magnifying Glass
47mSecrets of Performance Tuning Java on Kubernetes
27mPods on Steroids - A Case Study of Migrating Flipkart.com to K8s on Bare Metal
31mPodOverhead: Accounting for Greater Cluster Stability
15mStuffing 10K Pods on a Commodity Server - And Starting Them in Milliseconds via Kubernetes
40mJava Performance on Kubernetes