摘要： 由不同高性能RISC处理器及其构成的超级计算机系统，其基准性能和体系结构是不尽相同的，而且一种计算机体系结构的寿命也就只有3～5年时间，这将使得大多数付出巨大代价且在特定平台上开发成功的高性能科学计算程序往往会因为机型的改变而不能运行或不能很好运行，因此，人们迫切需要通用化程序，文中通过对在几种主流超级计算机上大量的测试数据的分析，指出了消息传递模式与指令级并行性对提高并行程序通用化程度的重要影响,同时也对利用网络化虚拟集群计算机资源从事高性能科学计算的可行性进行了讨论。

关键词: 高性能科学计算, 通用化, 消息传递模式, 网络化虚拟集群超级计算机, 指令级并行性

Abstract: A supercomputer system can be constructed by different high performance RISC processors and their benchmark performance and architecture are different accordingly. The lifetime of the architecture is usually only 3～5 years, which will lead to the failure of most high performance science computing programs developed at huge cost on the special computer because of the change of the machine type, so portable program is in urgent need. By analyzing a lot of testing data obtained on several main current supercomputers, this paper points out the important influence of massage passing mode and instruction-level parallelism on increasing parallel program's portability. At the same time, the paper also discusses the feasibility of utilizing networked virtual cluster supercomputer for high performance science computing.

中图分类号: 

• TP311.52