Research On IP Core Mapping Algorithms For Optical Network-on-Chip

With the continuous development of semiconductor technology,the number of processing cores integrated on a single processor is increasing.Network-on-Chip(NoC)has become the on-chip communication infrastructure for multi-core processors due to its good scalability and parallelism.Optical NoC(ONoC),which uses optical signals as the data transmission medium,has excellent performance such as high bandwidth,low latency,and low power consumption,make it a promising on-chip communication architecture.However,due to the inherent properties of optical signals and optical devices,optical signals suffer insertion loss and crosstalk noise when they transported in ONoC,which seriously affects the on-chip communication performance and multi-core processors' energy consumption.The insertion loss and crosstalk noise of ONoC are affected by the transmission path of optical signals,and application IP mapping result is related to the optical signals' paths on the ONoC,which affects the insertion loss and crosstalk noise in the ONoC.Therefore,how to perform IP core mapping on ONoC has become an important issue for researchers.This thesis proposes a single application mapping algorithm and a multi-application mapping algorithm for ONoC's IP core mapping problem.Aiming at ONoC's single application mapping,an NSGA-? based multi-objective IP-core mapping method is proposed in this thesis.It builds a multi-objective IP-core mapping mathematical model to optimize the insertion loss and crosstalk noise of ONoC at the same time.Moreover,an improved NSGA-? algorithm is applied to this model to obtain an appropriate mapping result.Experiment results show that,compared with the singleobjective algorithms,the proposed IP-core mapping method reduces insertion loss by average 15.5% with similar crosstalk noise,or reduces crosstalk noise by average 82.7%with similar insertion loss.Aiming at ONoC multi-application mapping,this thesis proposes a bin-based defragmentation algorithm DEBin for multi-application mapping in ONoCs.Considering the characteristics of ONoC based on optical routers Crux,in order to reduce the insertion noise and crosstalk loss of ONOC in multi-application mapping scenarios,and to improve the throughput of multi-core processors,the algorithm DEBin sorts the current running applications according to their size of the mapping area,and applications with same mapping areas' height are placed in the same bin,so that idle processing cores are gathered,so as to reserve the mapping area for subsequent applications.Experiment results show that using DEBin to defragment in multi-application mapping can increase the throughput of multicore processors by most 13.8%,while greatly reducing the waiting time of applications.