Timely Cluster System Design And Optimization

Based on open system standard, cluster computing system constructed by universal business products has already been adopted in high capability computing field abundantly. There are 360 cluster system structures out of world top 500 super computer systems, which is 72 percent of the top500's total number[1].Cluster system emerges in the background of commodity components becoming increasingly high-capability, stable and credible, low-cost. People connect servers, work stations, even top grade microcomputers by high-speed Internet to substitute expensive mainframe computer and MPP, etc. Depending on its favorable expansibility and performance-quotation ratio, cluster system rapidly becomes preferred system structure to construct super computer. Meanwhile, in virtue of huge data quantity transaction in types of application systems, the requirement of high density data throughput propose the higher demand on the high expansibility of cluster system. Distributing cluster dispatch and allocate function into back-end serve node, can greatly depress the burthen of front-end node traditionally supplying dispatch and system load balance mechanism, eliminate the bottle-neck of system expanding function. But rigorous requirement of response time limit in real time application also brings forward challenges to cluster system functions. Routine real time operation system does real time dispatch by kernel, centralizes computing resource to ensure the time limit of real time task. But the resource dispatch bases on stand-alone system, lacks of the function of supporting multi-node cluster system, and is often used in embed system, which has limited EMS memory, hard to support cluster structure. Currently, general method, adding cluster load balance scheduler in middle level, emphasizes particularly on load distributing in balance, which goal is to maximize system's output, but is hard to satisfy response time limit demand for real time task. Thus, researching real time dispatch mechanism for open cluster system, realizing partition of different time limit tasks, focusing on throughput demand while assuring time limit of real time task, have already become quite hot.The dissertation has researched actuality about real time system and cluster system,analysed real time cluster system's particular requirement.The following is detailed work: ①The dissertation proposes three real time scheduling strategies, does particular design on real time cluster system on Linux system platform, and constructs open real time cluster system.②The dissertation specially researches strategie replying on system overload situation, proposes task discarding mechanism to alleviate system pressure and prevent deterioration of system performance in overload situation.③After analyzing the developing trend of current system node and adequately considering correspond connection between CPU and IO in Linux operation system, the dissertation does researches on sequence serve pattern and subsequent serve pattern, meanwhile, proposes a flexibility control mechanism in subsequent pattern, to make multi-tasks can share system resource more reasonably in subsequent pattern.④Optimizes IO communication of current RTCS system, in the instance of simple sequence service mode, useing multi-IO method supplied by open platform to improve system response capability, achieve optimizing goal. Every above optimizing design is validated by experiment, the performances in throughput, response time, and discarding ratio all precede original RTCS cluster system.