Research And Application Of High-performance Pc Cluster

As the society evolves, the needs of High Performance Computing (HPC) are growing fast. The application of HPC is no longer restricted in certain areas, such as military or nuclear physics, it could be used in all areas of the society. With the developing of hardware and software technologies, the Clusters, which are assembled by commercial off-the-shelf (COTS) PCs with high-speed network and freely distributed software, deliver superior performance. As the cost of PC hardware and network equipment continues to rapidly descend, the Clusters have been accepted by more and more customers and it could be reflected in the "Top 500 Supercomputing Sites" ?a list that names the most powerful computers in the world.However, the COTS hardware still has some disadvantages compared with server-specific hardware, so we have to design optimal Cluster architecture and Cluster software system to make Cluster exhibit its high performance, high availability and scalability. The Cluster architecture is usually designed according to the application and the implementation of software system is based on the architecture. The Cluster system's performance and reliability depend on a range of key software modules, including communication module, load balance module and high availability module.This thesis introduces two examples of Cluster system's design and implementation, one is a Cluster Video Server for DVB system and the other is an Index Calculation System for decision-making. Either Video Server or Index Calculation System has strong requirements of computation performance and I/O throughput, besides that the systems must be reliable. These two kinds of servers are traditionally implemented by SMP servers, which are quite expensive. However, the systems could provide cost-effective services by using Clusters.The thesis gives the solutions to.implement the key software modules of Clusters by these two sample systems. It uses socket communication combined with thread technology or MPI functions directly in the communication module, which guarantee the communication efficiency. It also uses data strip method to implement the system's load balance, and the system performance wouldn't descend by careful memory management. Taking advantage of redundant equipment and carefully arranging the tasks' transition, the high availability problem is solved.Various tests have been taken on two systems and the result shows that the Cluster systems could meet the services' requirements.