Strategies for Mitigating Performance Interference in Cloud-Native Systems
summary
Claims that cache and memory contention cause measurable slowdowns: typical services show 20–30% slowdowns with some peaking at 3×, and Google reports cases of 5× and 14× regression. Detection uses CPI and a high‑resolution 1 millisecond granularity measurement pipeline implemented with eBPF and a Rust collector (collector uses about 350 MB and ~200 millicore CPU under load). Open source under Apache 2 with Helm charts; automatic throttling or core pinning are demonstrated mitigation paths.
AI-written summary · about
chapters (11)
- 0:00Performance interference intro
- 3:36Cache and memory noisy neighbors
- 7:50Noisy neighbors impact
- 10:45Why noisy neighbors matter
- 13:35Manual mitigation strategies
- 14:59Automatic CPU/memory controls
- 17:52High-precision observability at 1ms
- 20:42Detecting noisy neighbors
- 21:48Performance overhead & deployment
- 23:43Open-source tools & community
- 24:46Summary & Q&A
related talks
28mExtending Container Performance Isolation: Regulating Memory Bandwidth & Cache in...
15mOptimizing Application Performance on Kubernetes
49mNoisy Neighbor Detection: A New OSS Collector
31mRunning Low-Latency Workloads on Kubernetes
42mWhen the OS gets in the way
41mThe Cloud Native Compiler: JIT-as-a-Service
47mSecrets of Performance Tuning Java on Kubernetes
35mCPU Utilization: The Hidden Cost of Running Hot