Extracting method of packet dependence from NoC simulation traces using association rule mining

Abstract

Networks on-chip (NoCs) interconnect complex parallel applications on multiprocessors systems-on-chip. In order to rapidly evaluate NoCs, designers replace processing elements by communication traces on simulations. However, trace-based simulations have low accuracy due to the lack of information about packet dependence. The methods to obtain packet dependence require running multiple simulations or modifying application and simulator source code to output dependence; both are very costly. In this paper, we model packet dependence extraction as an association rule mining problem. We use the Apriori algorithm to extract communication patterns using traces from only one full-system simulation. The experiments with real, synthetic, and self-similar traffic show about 78% of packet dependence accuracy.