|
With the development of economy and science,the computational tasks become more and more in the fields of biomedicine,weather forecasting and high-power physics.Their specialties are:lots of computational data,complicated algorithm,large-scale,most difficult computation. So,Computer’s reliability and availability are becoming more and more important.Though a single computer’s capability and reliability are becoming better and better,But a single computer can’t complete some requests still.Therefore,people think of a method to build a high performance system with common PCs which are connected to each other by some ways,the system is called cluster.Cluster is the best substitute of expensive super computer,because it is very cheap and convenient,moreover,it demands very low for software and hardware.Cluster technology is a technology of organizing several computers to work together to make a more powerful computer.The technology can provide more availability and more reliability.As the HPC’s a branch,The application of Beowulf cluster system based on the Message Passing is getting more and more widespread,it has many merits,for example:inexpensive,easy to manage,the higher performance-to-price ratio.In this paper,we design and implement a cluster system based on Linux and MPI with existing hardware and software in the laboratory by introducing framework technology of cluster middleware.Firstly,in this paper,we discuss structure and function of the Linux cluster in detail,analyze the research situation of cluster in existence,the program model of parallel computing and Message Passing Mechanism.We analyze the basic principle,idea,composition and the main component function of the cluster system based on MPI and Linux,and research some key technologies about parallel computing and cluster middleware.Secondly,in this system,we design and implement a cluster middleware,including computing node insert module,resource management module,and task schedule module.By implementing function of some modules in cluster system,this system has functions of testing network load,dynamically reflect the change feature for node and rationally distribute task according to the load of node.Nowadays,these researches and implementations are hot research subjects on parallel computing system,and provide some references for relative researches.At the end of this thesis,in the experiment,a parallel program is used based on this cluster system to test the availability and robustness of the cluster system by using static dispatching algorithm and task schedule algorithm which we propose in the paper,separately.The experiment indicates that the cluster system has excellent computing performance and rapid response to the external performance,and it can dynamically reflect the change feature of capability parameter in the cluster.It has some good characteristics such as simplification,easy extensibility and so on.These researches and implementations in the paper are a hot research subject on parallel computing system.They shed light on some new ideas and plans for relative research.
|