Towards performance prediction of multicore software routers

Rapid advances in network function virtualization technologies have led to the emergence of flexible and scalable embedded data plane processing functions, eg, packet forwarding, on commodity hardware. However, performance prediction of softwarized network functions on such shared resources is challenging but very important for obtaining the full benefits of network function virtualization. This paper addresses the problem of performance prediction for multicore software routers and reveals a key technique to achieve high accuracy. Motivated by observations, we first analyze how many CPU cycles are spent for forwarding a packet in multicore processing systems. Our prediction model of the CPU usage based on cache contention can capture its nonlinear dilation scaled by the number of CPU cores–called dilated CPU consumption (DCC). We validate the accuracy of the DCC model with measured data. On the basis of the DCC model, we develop 2 performance prediction algorithms to predict the maximum throughput of the packet‐forwarding function corresponding to the assigned resources. The first algorithm includes CPU utilization statistics (called DCC‐u), while its simplified version (called sDCC) does not require CPU utilization statistics. We validate the proposed models through the exhaustive set of experiments under various amounts of assigned resources (ie, CPU speeds and the number of CPU cores) and different traffic loads (ie, a large and small number of traffic flows). The results show that DCC‐u and sDCC achieve high accuracy under various amounts of resources and traffic conditions. Remarkably, they improve the precision of estimation from that of the existing techniques by up to 105% to 163%.