Access network traffic models for energy efficient architectures

In this paper we propose a novel technique to infer models of access network traffic from real-world traces and use them to evaluate the energy efficiency of microarchitectures for network devices. Unlike static packet traces, stochastic models of network traffic enable the simulation of reactive behaviors and are vital to the prediction of energy efficiency in network devices. While it is widely accepted that Internet backbone traffic can be represented by self-similar Markov-modulated Poisson processes, traffic patterns in access networks have been found to escape this characterization. We show that accurate models for access traffic can be automatically constructed employing causal state-splitting reconstruction on an alphabet obtained by k-means clustering of the characteristics of the sampled traffic. Applications concurring for the same bandwidth generate aggregated flows: each flow is only loosely coupled to the others, but every flow bears individually a strong sequential correlation. We infer a model for every individual flow, after we have separated it employing a suite for deep packet inspection. These models generate parallel traffic flows whose aggregation is reconstructed by a statistical shaper. Our cycle accurate simulations demonstrate that the energy efficiency predicted by our traffic models for dual and quad cores versions of a multicore packet processing microarchitecture bears consistently an absolute error lower than 2% against values obtained on real traffic.