Extending Container Performance Isolation: Regulating Memory Bandwidth & Cache in...
summary
Claim: shared last-level caches and memory bandwidth can slow services up to 4–13x and produce millisecond-scale spikes; existing resctrl + cgroup filesystem is too slow and misses new containers. Solution: a collector is built that captures per-process millisecond measurements and enforces limits using eBPF + perf + hrtimers, avoiding expensive syscalls and file parsing. Demo on the OpenTelemetry app shows 20–30% typical slowdowns and short 3x stalls, and the collector successfully detects and can throttle offenders.
AI-written summary · about
chapters (9)
- 0:00Performance isolation intro
- 3:14Noisy neighbor problem
- 8:13Memory contention impact
- 12:54Linux isolation mechanisms
- 15:05Limitations of controls
- 20:27Kernel improvements proposed
- 23:20Isolation for other resources
- 25:31Open source tools & next steps
- 26:38Conclusion & Q&A
related talks
26mStrategies for Mitigating Performance Interference in Cloud-Native Systems
36mMitigating Noisy Neighbours: Advanced Container Resource Management
13mHarvesting Spare CPU Resources in Container Systems
34mContainer Runtimes... on Lockdown: The Hidden Costs of Multi-tenant Workloads
32mContainer Escape Paths Nobody Monitors: Abusing Linux Debug Interfaces
31mYour Containers Aren’t Alone: Demystifying Container Isolation
30mRestricted Address Spaces for Container Security
47mSecrets of Performance Tuning Java on Kubernetes