The Research And Optimization Of Inter-core Communication In Multicore Architecture

The emergence of multicore promoted the rapid development of related technologies, however, problems such as communication delay, memory wall and resources contention affect the performance of multicore processors. Improving the efficiency of multicore communication is an important way to improve system performance, and this has become the hot spot in field of computer architecture.The premise of inter-core communication is to ensure applications run normally. The nature of the research is to coordinate the applications execution order and distribute applications to multicores, so as to reduce or even avoid shared resources contention, and ultimately improve multicore system performance.Currently, the industry has proposed several strategies to improve inter-core communication, including the following three aspects:1.Improve MPI (Message Passing Interface) library;2.1mprove specific network(such as mesh network);3.Propose more efficient process mapping algorithms to reduce inter-core communication overhead. Considering the versatility and feasibility, this paper studied from the third point and put forward two process schedule algorithms for two different multicore platforms.(1) This paper proposed a schedule algorithm which can schedule a few tasks in advance by one cycle for periodic tasks on embedded multicore platform. By introducing pipeline, the algorithm found out the communication task TRmx which occupying bus with the longest time and all computing tasks which relied on TRmx, and the algorithm schedule all the tasks in advance from the beginning computing task T1to the communication task TRmx, and then in periods afterwards the data needed is ready before the tasks which relied TRmx start to run. As a result, computing tasks and communication tasks can run in parallel, achieving the pipeline effect. Tasks between different periods can execute in parallel in order to shorten schedule length.(2) This paper established inter-core communication time model on NUMA architecture and studied several factors related to inter-core communication efficiency, such as message length、message rate、network interface utilization etc. By experiments the paper quantified the influence degree of every factor for inter-core communication time, and then set a threshold for the number of processors which can run in a node based on the limited network interface contention. In conclusion, the paper concluded several basic rules to apply in the process mapping algorithm.By multi-group comparison tests and in-depth data analysis, this paper concluded above two algorithms both can achieve the aim to improve multi-core processor system performance compared to traditional schedule algorithms.