Study on Interconnection Networks that Employ Autonomous Optimization Functions based on Global Behavior Understanding

点击翻译按钮获取中文摘要

This research mainly focuses on interconnection networks for massively parallel computing systems. Its major aim is to enable an optimized control of the whole interconnection network by introducing autonomous and collective operations of network routers. This research is based two major viewpoints ; processor behavior (i.e., program execution property) and global, perspective understanding of network.The former viewpoint progresses the 'meta-level computation' concept that is one of the central outcomes of the prior research. The results include (1) profiling methods, from both of software/hardware viewpoints, (2) basic effects of optimization techniques, (3) multithreading methods, including speculative execution.From the latter viewpoint, we complete (4) Cross-Line interconnection network routing method that uses quasi-global information of congestion and (5) Eagle fault-tolerant computational grid system. We extend the Cross-Line method to fault-tolerant features and (6) propose a new method Fault-Tolerant Cross-line. Furthermore, we extend the core concept of autonomous operation to apply P2P (peer-to-peer) system.From deep discussions on behaviors of program execution and network operation, (8) we reach statistics concepts based on statistical mechanics (in physics) and information theory. The common term is 'entropy. Information entropy is effectively used for program behavior and properties of some processor components. Thermodynamics entropy is useful for aggregates of network packets. These entropies are useful for numerical representation of behaviors. This idea is very fruitful and we continue in the successive research.