Measuring K8s Network Performance
Talkat KubeCon + CloudNativeCon North America 2021 · on Friday, 2021-10-15· at 16:30· Room 502 AB + Online·
summary
Benchmarking Kubernetes networking requires a precise question, controlled setup and repeated runs (≥3) with full hardware and software snapshots recorded. Use netperf or the cluster-focused cubenet-bench to run point‑to‑point, echo/RTT and multi‑stream tests; expect a single TCP stream to hit ~20 Gbps on one core so saturating 100Gb requires multiple streams and pinned CPUs. An unexpected result—native routing slower than tunneling—traced to veth handling and is fixed in newer kernels; isolate experiments by running bare metal, disabling preemption and pin NIC interrupts to avoid noise.
AI-written summary · about
chapters (15)
- 0:00Who am I? What do I do?
- 0:46Why measure K8s net performance?
- 2:43The stack
- 5:51Evaluation process
- 6:49Evaluation goal
- 10:33STREAM
- 14:01ECHO-REPLY: N reqs in-flight
- 16:47Scaling workloads
- 18:47Metrics
- 20:55System configuration
- 22:21Experiment
- 23:39Cilium war story
- 25:21Tooling
- 27:06Resources
- 27:59Conclusion
related talks
35mDeep Dive: Kubernetes (Instrumentation) SIG
06mBenchmarking Journey Through Service Meshes
35mWho Dropped the Packet? Solving K8s Network Mysteries in Real-Time
31mA High-Schooler’s Guide to Kubernetes Network Observability
27mA Service Mesh Benchmark You Can Trust
18mMethodology and Practice in Observing Kernel Networking
36mUnderstanding Scalability and Performance in the Kubernetes Master
30mLive Experiments with K8s Applications: Pitfalls and Ho...