Traditional NFVi dataplane performance benchmarking and testing rarely focuses on availability and consumption of resources such as CPU Cache and Memory Bandwidth. Quality of Service fine tuning, without considering such resources, often leads to problems such as over-subscription and shared resource contention - often termed as noisy neighbor effects. This work focuses on mitigating noisy-neighbor effects with the efficient management of such resources.
We present the methodology and test scenarios using opensource solutions - OPNFV-VSPERF, Collectd, Stress-NG, and RMD-client - for (a) analyzing the performance impact due to the presence of noisy neighbors and (b) controlling access to Last Level Cache (LLC) to achieve optimal performance. The presentation will cover the creation and analysis of LLC profiles of all workloads, and policy-driven management of LLC resource to improve the performance and manage the noisy neighbors.
