Research And Implementation Of Small Linux-Based Cluster

Traditional single-processor computing can’t meet computing requirements, and at the sametime, the high-performance computer based on vector processor and symmetric multi-processorshas many problems, such as expensive, low scalability, poor expansibility. In order to accelerate thespeed of solving the problem, the paper builds a cluster system under Liunx environment.Thesystem has good scalability, availability and cost-effective.A Linux-based high-performance cluster includes many software modules, such as NFS filesharing system, MPI parallel programming library and so on. And then the paper compares theserial and parallel algorithms, which calculate the value of PI by using the C+MPI programmingmode.The numerical results demonstrate the feasibility of parallel computing and the advantages ofparallel algorithm.In order to solve the degree of parallelism caused by different matrix division method, thepaper uses two methods (the striped matrix multiplication parallel algorithm and the checkerboarddivided matrix multiplication parallel algorithm) to analyze the computation time and otherperformance indicators. The experimental results show that the checkerboard division method candevelop a higher degree of parallelism than the striped method.The thesis analyzes the parameters of HPL test which includes the dimension of the coefficientmatrix in solving dense linear equations, the LU decomposition block size, processor mesh size, thenumber of processors participating in the calculation and other parameters of HPL package. Thepaper evaluates the floating point performance of cluster system by adjusting the parameters inHPL.dat. And the Optimal floating point performance of the cluster system is13.61Gflops.