Noisy Neighbor Detection: A New OSS Collector
summary
Latency can jump 4–13x from memory contention and garbage collection, and second‑scale collectors miss it. The project builds an OSS collector that samples per‑process memory metrics at millisecond sampling across cores using eBPF and CPU perf counters, writes 1 ms traces to Parquet, and currently reports few‑hundred MB memory and ~250 milli‑cores overhead on a 30‑core workload. Detection works in demos and enables attribution and automated throttling or isolation.
AI-written summary · about
chapters (7)
- 0:00Noisy neighbor problem
- 3:15Problem importance
- 13:52Measuring and mitigating noisy neighbors
- 18:05Millisecond-granularity collector
- 26:09Deployment and overhead
- 27:33Community and future steps
- 28:15Q&A
related talks
49mNoisy Neighbor Detection: A New OSS Collector
34mThe Missing Metrics: Measuring Memory Interference in Cloud Native Systems
18mNoisy Neighbor Detection with eBPF
23mControlling Data Overhead with the OpenTelemetry Collector
25mMaking Observability Work at the Edge
15mTuning OTel Collector Performance Through Profiling
30mNoisy Neighbors, through Networking
57mTelemetry Showdown: Fluent Bit vs. OpenTelemetry Collector - A Comprehensive Benchmark Analysis