Optimization Exploration Of Low Energy Mapping For Irregular 3D NoCs

With the improvement of system integration and development of Network-on-Chip technology,the three dimensional Network-on-Chip(3D No C)has become research focus of national and international academia and industrial with its features of stronger performance,lower latency,and higher scalability.Through-Silicon-Via(TSV)is employed for inter-layer connectivity(vertical channel)and is considered a solution with sufficient potential.It is an important direction to study the 3D No C structure by effectively reducing the number of TSVs used,however,because of its high cost and low yield.The mapping problem refers to how to place each application task on a 3D No C processing element to achieve a better target under satisfying constraints,which has a great influence on the performance and power consumption of the system.This thesis focuses on optimization exploration of low energy mapping for irregular 3D No Cs based on TSV technology.This thesis studies the structural characteristics of 3D No C with two kinds of irregular topology structures,and discusses the extensive research of homogeneous or heterogeneous processing elements.The simulation platform 3DNo CMap is customized and implemented for structural characteristics of irregular 3D No C with the capacity of static and dynamic mapping of 3D No C with irregular topologies based on researching No CMap and other simulation platforms,due to the vast majority of simulation platforms can't support mapping simulation of 3D No C with irregular topological structures.Meanwhile,in the 3DNo CMap,the routing algorithm is improved according to the characteristics of irregular 3D No C,and the power consumption model is established,the simulation platform is verified at last.In this thesis,the static mapping model is established.We analyze the principles and disadvantages of several typical heuristic mapping algorithms,and obtain ideas for improvement.Based on this,the hyper-heuristic algorithm is studied and applied to mapping problem,we design the hyper-helical mapping algorithm based on online learning which provides a richer search method for mapping schemes and online adjustment capabilities.Simulation results show that our proposed algorithm can effectively reduce the communication power consumption of No Cs,and decrease routing hop count and average link load compared with other algorithms.This thesis establishes a dynamic mapping model,explains and compares the current mainstream dynamic mapping algorithms,and analyzes three effective dynamic mapping strategies: packet ordering strategy,communication aware strategy,and sorting mapping strategy and make valuable improvements to the 3D No C with irregular topology.We propose communication energy aware and packing order mapping algorithm based on the advantages of these strategies.Simulation results show that,it performs better than other algorithms in terms of communication power consumption,hop count,and average link load with increasing tolerable algorithm execution time.