Multifractal on-chip traffic generation under TLM

In the design flow of Multi-Processed Systems-on-Chip (MPSoCs), the evaluation of communication structures play a very important role, since it may reveal relevant information on performance, energy consumption and cost. Simulation under a number of stimulus given by a traffic generator is a relevant solution for MPSoCs performance analysis. Traditional traffic generators based on Poisson and classic Markovian models are not able to reproduce certain characteristics of original application traces, such as bursts and self-similarity. After the detection of Long Range Dependence (LRD) in on-chip traffic, monofractal models started being used for traffic generation. These models, however, were mainly used under RTL/CA simulations and present statistical limitations, opening opportunities for tests with multifractal models and higher abstraction level descriptions. In this work, it is shown that the Multifractal Wavelet Model (MWM) presents a better accuracy in the modeling of on-chip traffic when compared with auto-regressive (monofractal) models and that the usage of traffic generators modeled under TLM can achieve simulation speed-ups in the order of 12x.