Abstract

The performance evaluation of supercomputers is usually done through a particular benchmark set. Today, we have a variety type of supercomputers, such as multi-vector-piped, SIMD-parallel, MIMD-parallel and their combination. For machines with such new architectures, what becomes much more important to achieve higher performance is the applicability of a particular program to the new hardware feature. The authors figure out that as doing performance evaluation there are two complementary objectives, one for users and the other for computer architects. Hence, it becomes a new goal of performance evaluation of a supercomputer that the benchmark programs should characterize both the behavior of attached special hardware and the structure of a user's program, then the results of benchmarking should be fed back to improve the architecture. Recently, in order to make the performance evaluation of new supercomputers, Newborn benchmarks have been proposed to meet new architecture. In this article, recent benchmarking trend and some New-born benchmarks will be briefly overviewed.