The Missing Metrics: Measuring Memory Interference in Cloud Native Systems
Talkat KubeCon + CloudNativeCon Europe 2025 · on Friday, 2025-04-04· at 11:00· Level 1 | Hall Entrance N10 | Room G·
summary
Cloud-native nodes suffer cache eviction and memory-bandwidth contention that repeatedly inflates P95/P99 by 4–14x and cuts CPU efficiency by ~25%; major operators (Google, Alibaba, Meta) observe this in production. The project measures per-core memory and cache usage with 1 millisecond sampling and RMID tagging, writes raw telemetry to Parquet (roughly ~1 MB/s raw telemetry), and aims for 0.1% overhead. Detection enables actions via resource control (pinning, frequency limits, anti-affinity or automatic throttling) and has reduced tail latency in examples.
AI-written summary · about
chapters (11)
- 0:00Memory noisy neighbors intro
- 4:39What is memory noisy neighbor
- 11:44Why interference matters
- 15:56Mitigating noisy neighbors benefits
- 19:11Measuring and mitigating interference
- 23:09Building Unvariance collector
- 26:22Call to action overview
- 27:03Q&A: Cache vs bandwidth
- 29:37Q&A: Storage for sampling
- 31:03Q&A: Continuous monitoring
- 32:06Q&A: Detecting noisy neighbors
related talks
49mNoisy Neighbor Detection: A New OSS Collector
21mZero Overhead Monitoring for Cloud-native Infrastructure using RDMA
34mNoisy Neighbor Detection: A New OSS Collector
36mMemory Why You Need to Care
16mMicroMon: A Monitoring Framework for Tackling Distributed Heterogeneity
36mMetrics - What's your code actually doing?
18mResource Guarantees in a Multi-Tenant Infrastructure
47mResolving performance issues in practice