Performance analysis of resource pooling for network function virtualization

In the framework of network function virtualization, we consider in this paper the execution of Virtualized Network Functions (VNFs) in data centers whose computing capacities are limited. We assume that each VNF is composed of sub-functions to be executed on general purpose hardware, each sub-function requiring a random amount of processing time. Because of limited processing capacity, we investigate the relevance of resource pooling where available cores in a data center are shared by active VNFs. We study by simulation various algorithms for scheduling sub-functions composing active VNFs (namely Greedy, Round Robin and Dedicated Core algorithms). We additionally introduce an execution deadline criterion, which means that VNFs can renege if their sojourn time in the system exceeds by a certain factor their service time. This feature is especially relevant when considering the processing of real-time VNF. Simulations show that sub-functions chaining is critical with regard to performance. When sub-functions have to be executed in series, the simple Dedicated Core algorithm is the most efficient. When sub-functions can be executed in parallel, Greedy or Round Robin algorithms offer similar performance and outperform the Dedicated Core algorithm. Enabling as much as possible parallelism and avoiding chaining when designing a VNF are fundamental principles to gain from the available computing resources.