3D-AMAP: A Latency-Aware Task Mapping onto 3D Mesh-Based NoCs with Partially-Filled TSVs

摘要

This paper proposes a latency-aware task mapping algorithm called 3D-AMAP for 3D mesh-based NoCs with partially-filled TSVs. The 3D-AMAP algorithm divides communications of a given application graph into Low-volume (LV) and High-Volume (HV) communications. The 3D-AMAP algorithm bypasses the LV communications to partition the given application graph to some subgraphs. Then, 3D-AMAP algorithm fairly assigns 4-neighbor cores of the mesh topology between the high traffic rate tasks of the application graph to reach the bestmapping. The proposed mapping algorithm maps application subgraphs one by one based on their total intra communications considering where the vertical channels are located in the network. Evaluations of the 3D-AMAP mapping algorithm are done in a wide range of working conditions using Access Noxim NoC simulator in terms of network latency. Results show that 3D-AMAP algorithm offers at least 5% and at most 76% in network latencywith respect to NMAP algorithm.