Guiltiness: A Practical Approach for Quantifying Virtual Network Functions Performance

Abstract In Network Functions Virtualization (NFV), service providers create customized network services by chaining Virtual Network Functions (VNFs) in forwarding graphs, according to individual client demands. Despite the flexibility promoted by the NFV paradigm, specific VNFs used in network services may become bottlenecks due to a number of factors, e.g. , lack of resource capacities, demand overload, incorrect ordering, and interdependency between VNFs. Hence, resource monitoring is crucial to determine which VNFs have a negative impact on the quality of service. However, NFV imposes constraints that hinder the adoption of traditional network monitoring approaches, such as: the heterogeneity of environments with all sorts of VNF purposes ( e.g. , caching, address translation, and performance optimization), and the atypical functioning of specific VNFs that rely on non-blocking I/O implementations. In this paper, we propose a model to quantify the guiltiness of each VNF on degrading the performance of a network service. This model consists of a novel application of the Utilization Law from queuing networks theory. Through numeric assessments on typical scenarios, we refined the set of relevant resource metrics and applied a weighted sum to gauge them in the model. Also, a hybrid algorithm based on linear regression and neural networks is introduced to adjust the model’s parameters according to the environment particularities, such as the type and number of VNFs in the service. Experimental evaluations confirm the ability of the model to ( i )detect bottlenecks, and ( ii )quantify performance degradations. When capacity restrictions were applied to different types of VNFs, our guiltiness metric was able to detect the root cause of degradations. Further, the guiltiness metric outperformed traditional metrics, being able to characterize 96.15% of the performance issues with a 25.6% reduction in the number of false positives when compared to CPU usage.