Noisy Neighbor Detection: A New OSS Collector
summary
An open source collector samples every core at 1 millisecond sampling using eBPF + Perf counters and high-resolution timers to compute CPI cycles per instruction and detect memory contention that raises tail latency (Google shows up to 5x, mem-style stores up to 14x). The collector writes raw millisecond parquet traces to S3, keeps under ~350 MB RAM and ~0.75% per-process CPU overhead, detects services noisy a few percent of the time, and enables offline detectors or automated core/jail mitigations.
AI-written summary · about
chapters (8)
- 0:00Noisy neighbor problem
- 4:16Memory noisy neighbor
- 14:55Care about memory noisy neighbors
- 17:30Measure and mitigate noisy neighbors
- 23:321ms collector for detection
- 31:03Open questions and future work
- 35:42Deploying OSS project
- 36:23Q&A
related talks
34mNoisy Neighbor Detection: A New OSS Collector
18mNoisy Neighbor Detection with eBPF
26mStrategies for Mitigating Performance Interference in Cloud-Native Systems
34mThe Missing Metrics: Measuring Memory Interference in Cloud Native Systems
25mMaking Observability Work at the Edge
23mControlling Data Overhead with the OpenTelemetry Collector
01h 13mContinuous Profiling: The Missing Piece in Your Observability Puzzle
30mNoisy Neighbors, through Networking