Extending Container Performance Isolation: Regulating Memory Bandwidth & Cache in the Kernel
summary
Shows that shared cache and memory bandwidth cause real slowdowns — measured examples range from 3–14x latency increases and common apps suffer 20–30% CPI degradation during contention. Describes why resctrl’s filesystem API cannot measure fast-changing behavior (global lock, coarse sampling) and presents a collector that records per-process cache/bandwidth at millisecond sampling using eBPF + perf + high-resolution timers. Proposes adding perf-backed resctrl support and explains RMID limbo handling and practical tradeoffs.
AI-written summary · about
chapters (7)
- 0:00Memory noisy neighbor problem
- 3:17Memory bandwidth and cache contention
- 11:51Linux kernel resource control
- 17:37Limitations of current controls
- 26:17Perf subsystem integration proposal
- 34:45Memory contention impacts and solutions
- 38:47Q&A
related talks
42mWhen the OS gets in the way
36mMitigating Noisy Neighbours: Advanced Container Resource Management
21mDo Not Disturb Mode: Fine-Tuning Resources for Latency Sensitive Workloads
37mAdvanced Performance Profiling Using Perf Tools
31mMeasuring the Impacts of the Preempt-RT Patch
01h 05mPerformance Optimization in Software Development - Being Friendly to Your Hardware
39mThrottling: New Developments in Application Performance with CPU Limits
19mP2CACHE: Exploring Tiered Memory for In-Kernel File Systems Caching