Based On The On-chip Self-similarity Of Network Traffic Analysis And Modeling

Networks-on-Chip (NoC) was proposed as a state-of-the-art communication architecture of System-on-Chip (SoC) design. Traditional bus-base SoC is confronted with problems such as inefficient communication, poor scalability and limited integration level, when deep-sub-micron CMOS technology is adopted. Network-based Multi-Processor SoC (MPSoC) enjoys the benefit of unlimited scalability, flexible clock trees and convenient floor-planning due to the Globally Asynchronous Locally Synchronous (GALS) clock infrastructure. Packet-switching-based inter-core communication provides efficient and low latency interconnection. Standardized infrastructure and regular topology make the Intellectual Property Reuse methodology more versatile in NoC-based MPSoC.The characteristic of network traffic has great impact on network performance. Internet Traffic engineering has got substantial research achievements and facilitates the development of Internet profoundly. As NoC technology thriving in recent years, researchers realized that the traffic of NoC should not imitated by traditional Short Range Dependence (SRD) traffic model such as Poisson and Markov process, and the existence of Long Range Dependence (LRD) property in specific applications, Mpeg2 decoder for example, has been validated. Therefore, using Self-similarity theory to analysis NoC traffic, building up a LRD traffic model according to actual traffic, and the synthesis of injection traffic for simulation are new exploration insights of NoC design.Taking the Internet traffic modeling theory as reference and combining it with the special property of NoC, this paper verified the existence of Self-similar characteristics of NoC traffic by experiments firstly, then calculated Hurst parameters of traffic obtained from OFDM baseband NoC routers. After analysis of the potential causes of NoC traffic Self-similarity, the variation of Hurst parameter according to processor pipelined degree and the number of virtual channel has been researched under a typical traffic distribution model. Finally, MMPP(2) method is adopted to synthesis the self-similar traffic, which is injected in simulator and compared with Poisson injection traffic. Simulation results show that at low injection rate, self-similar traffic deteriorates NoC performance, but at high injection rate exceeding NoC's bandwidth, self-similar traffic's impact is inapparent. Thus using self-similar traffic to verify the NoC design can give a more reliable solution and avoid the design risk when type-out.