大规模并行应用程序的可扩展性研究

RESEARCH ON THE SCALABILITY OF THE LARGE SCALE PARALLEL APPLICATION PROGRAMS

为适应未来超大型并行计算 ,要求算法和应用程序必须具有良好的可扩展性 .以往的可扩展性研究更强调于对算法的分析 ,而对于实际程序可扩展性低的原因很少进行深入探讨 ,不能有针对性地指导用户改进程序 .现提出了数值可扩展性和并行可扩展性 ,用来描述并行系统的数值性能和并行性能的扩展行为 .并深入地讨论了数值可扩展性和并行可扩展性可能低的原因 ,提出了一套可扩展性评价准则 .使用这套评价准则和近优可扩展性方法 ,对一个大规模应用程序——二维等离子体粒子云网格法并行程序进行了分析 ,结果表明这套可扩展性评价准则可以帮助定位引起可扩展性低的原因 ,同时也表明 ,对于实际的大规模应用 ,在已知小规模问题的执行信息下 ,近优可扩展性分析方法提供了一种预测更大规模的问题在多少台处理机上运行更合理的途径 .这里的“合理”。

Future supercomputing demands that large scale parallel algorithms and applications have good scalability. Previous scalability studies lay stress on the studies of the algorithms scalability,but few on that of the application programs. They couldn't give users the information about how to adjust programs to improve its performance. The numerical scalability and parallel scalability are provided to describe whether the parallel system maintains its numerical attributes and parallel attributes. Furthermore, a suit of scalability evaluation criterion is provided to help the user to find the reason causing the bad scalability and to modify programs. This criterion and the near optimal scalability method are used to analyze the scalability of a large scale application program, namely two\|dimensional electromagnetic plasma with particle in cell method. Results show that the criteria help to locate the reason why the scalability is bad, and that the near optimal scalability method provides an approach to predict how many processors are to be used by a larger problem to get a reasonable utility, where its time is near to the shortest time to run and its efficiency is much improved.