Research On 2D-mesh Topology Classification Based NoC Mapping Technology

Network-on-Chip (NoC) mapping, which maps the application tasks to realistic network resources, is regarded as the key step in NoC design and the important factor of network performance. Different NoC mapping results could make a great affection on communication power, convergence speed, reliability and temperature performance etc. And different mapping technologies have direct-action on mapping results. Currently, the research on mapping is confronted with two problems: firstly, in regular topological structures, symmetry of topology is considered inadequate by mapping algorithms which results in a low convergence speed; secondly, while NoC changes topological structure, i.e. from regular to irregular, traditional mapping algorithm will not work effectively, and this problem is more prominent due to immaturity of architecture-aware technology. The dissertation attempts to research NoC mapping technology based on 2D-mesh topology classification.After an in-depth analysis of existing mapping algorithms, a mapping algorithm for regular topology is put forward in the dissertation, and then extended to irregular topology, to establish a mapping model for architecture-aware, and a chaos genetic mapping algorithm is further proposed. The main subjects of the dissertation are as follows:1. A mapping algorithm (MARM) for regular topological structure of 2D-mesh is proposed. Considering the best of symmetrical features of 2D-mesh, we propose a mapping equivalent place concept, and use of the principles of branch and bound as well as"pruning"to eliminate effects of mapping equivalent place on mapping convergence efficiency. Simulation results indicate that the algorithm can save 41%~61% compared with random mapping algorithm and 4%～60% compared with GA in the aspects of communication power. The algorithm can also reach a higher convergence speed.2. An architecture-aware model (A3MAP-SRM) based on special reachable matrix is proposed, which introduces metric parameters for communications volume scale, and forms a special reachable parameter matrix with reachable matrix, to settle the problems in A3MAP model, i.e. excessive errors for function of total distortion and the blind treatment of inaccessible paths. Simulation results indicate that for regular topology, the model applies to communication-intensive task mapping of large and medium scales; for irregular topology the model will settle the problem of task mapping inaccessible for communication module, as well as realizes a 10% optimization effect.3. A chaos genetic algorithm (CGNM) for A3MAP-SRM model is proposed, based on the peculiarity that A3MAP-SRM mathematical model is a kind of 0-1 integer programming, and considering the"premature"of A3MAP-GA algorithm, which embeds chaotic search algorithm into genetic algorithm to produce a new unit and model, to enhance population diversity and search efficiency. Furthermore, power function carrier technology is also introduced to improve the traversal performance of solution space for chaotic search. Simulation results indicate that the algorithm is more advanced than A3MAP-GA in the aspects of communication power and convergence speed, and has a better performance in the A3MAP-SRM model.