A NoC-oriented Prediction Mechanism for Communication Behavior of Program

The advent of network on chip (NoC) provides a high bandwidth and scalable solution for communication problems on the chip, but multi-core communication is still the main factor restricting the performance improvement of the network on chip. This paper proposes a run length encoding (RLE) prediction mechanism based on trust value named RLE-N+CON. The mechanism takes the vector made up of communication traffic between different node pairs as the characterization of current communication behavior of the program. The vectors in sequence is clustered into a number of phases, and phase series represents the historical behavior. According to trust value, the mechanism choose last value prediction or table-based prediction which compress the storage of the same continuous phase sequences to predict which phase the program belongs to in the next time unit. By using the GEM5 simulator to verify, the result shows that the prediction mechanism proposed in this paper can accurately predict the communication behavior of parallel programs. Prediction results has important significance to guide the allocation of cyber source to improve network performance.