Troubleshooting Low Latency Application on CNF Deployment
summary
Containers cause measurable regressions: throughput drops (50 G → 48 G) and tail latency climbs (60 µs typical to 75–150 µs) when a second instance is added, and a vRAN pipeline shows mset cleanup jumps 15 µs → 120 µs under co-scheduled containers. An ERSPAN analyzer runs 25 G on bare metal but stalls NVMe writes when four container instances run. Observability is limited: perf/PEBS/IBS often blocked by container/kernel privileges and builds differ by compiler flags and libs. Remedies tried include core-pin init files, a dpdk CPU-layout patch and eBPF runtime hooks for per-thread telemetry.
AI-written summary · about
chapters (11)
- 0:00Introduction and agenda
- 2:06Common container deployment issues
- 8:05Use case 1 pipeline model scaling
- 9:29Use case 2 run to completion model
- 12:22Use case 3 vRAN latency problems
- 13:25Debugging challenges in containers
- 16:35Root causes of performance issues
- 20:09Hardware awareness in containers
- 23:04Deterministic latency requirements
- 26:37Troubleshooting and solutions
- 29:29Q&A session
related talks
15mMitigating Scalability Walls of RDMA-based Container Networks
34mComprehensive Performance Benchmark on Various Well-Known CNI Plugins
26mContainer Networking: The Play of BPF and Network NS with different...
24mDeploying Container Network Functions (CNF) on Kubernetes
25mScaling Up, One Network Bottleneck at a Time
41mLinux Systems Performance
18mNoisy Neighbor Detection with eBPF
23mNetwork Anomaly Event Prediction and Optimal Resource Control in Cloud Native Network Functions