Performance Analysis And Optimization Of Current Parallel Programming Models For Many-core Systems

With advancement in integrated circuit manufacturing process and promoting of demands, the number of single-chip processor core will be presented a sustained growth trend, future processor chips will integrate hundreds or even thousands or more processors cores which is often called many-core processors. Although parallel programming, parallel programming model and compiler technology has experienced decades of development history, compared with the rapid development of hardware technology, software technology slow progress. The development of parallel application and research of parallel programming model and compiler technology faced with severe challenge.On one hand, many-core parallel programming difficult parallel programming not only need to consider the module division, logical structure and process control for many-core processors compared with the traditional serial program, but also consider the data division, thread synchronization, data sharing and other issues. In the case of many problems to be solved for serial programming, it will become more difficult for many-core processors to achieve efficient parallel programming. Therefore, it is important to achieve maximize parallelism for many-core system by analyzing the performance of parallel programming model and reducing the complexity from serial applications transformation to parallel program for many-core.With the continuous development of processor technology and the increasingly large number of cores in the Many-core processors chip, it faces increasing challenges about how to promote application performance and optimize resource utilization. Specifically, the new structure is more complex computer systems which based on heterogeneous many-core and it is difficult to quickly locate problems and tune When performance problems. In addition, due to the particularity heterogeneous many-core system, software is responsible for explicitly managing a large number of hardware resources, which make system performance more dependent on software technology. Therefore, if we do not consider the characteristics of many-core processor systems, it will inevitably result in no fully using the performance advantages of many-core processors, and even ultimately affect system performance to run multiple applications or threads simultaneously in the multi-core.This paper has researched parallel programming model around many-core structure, and then explored solutions to improve the task scheduling strategy for parallel programming model in a systematic approach to many-core system architecture. Firstly, I introduced several typical many-core parallel programming model; Secondly, I analyzed their actual performance in detail by a large number of experiments. Finally, I proposed an adaptive scheduling framework A-SYS, which based on dynamic feedback control theory and integrated resource allocation, runtime control, task execution for solving parallel programming model scheduling problems in many-core system.