Abstract: In the distributed shared memory cluster system, this paper presents an OpenMP parallel computation method with explicit data distribution for nonlinear diffuse equation. The method partitions the whole data into small blocks, puts the block to the OpenMP thread and transfers data between the thread by using the copies in the shared memory. It designs a global reduction algorithm to reduce the synchronization times. Performance analysis and test results show that this method has well scalability on the distributed shared memory cluster with Intel Xeon quad-cores processors. Parallel speedup relative to 64 is 7.06 on 1 024 cores.

Key words: OpenMP programming model, multi-core cluster system, nonlinear diffuse equation, parallel computation

摘要： 在分布共享的多核集群系统中，提出一种求解非线性扩散方程的显式数据分布OpenMP并行计算方法。将数据进行分布式划分后分配到每个OpenMP线程，通过数据拷贝实现同步计算，并设计全局归约算法减少障碍同步次数。性能分析和测试结果表明，该方法在 4核Xeon处理器构成的分布共享集群系统上可扩展到1 024个CPU核，相对于64个CPU核，其加速比为7.06。

关键词: OpenMP编程模型, 多核集群系统, 非线性扩散方程, 并行计算

CLC Number: 

• TP301.6