The field of multimedia systems is becoming increasingly important in both the research and commercial worlds. Multimedia programs impose demanding and stringent requirements on all aspects of a computer system. However, there is a dearth of good benchmarks for such software, and there has been no rigorous performance comparison of different systems for such programs. In this paper, we describe a multimedia benchmark suite consisting of six programs and present the results of running these programs on five different systems. Our results show that in general, integer performance is not a good metric to use to predict multimedia performance. For these programs, performance on our DEC 5240 and HP 9000/730 had a high degree of correlation with SPECInt92, but varied between 48% and 118% for the DEC Alpha 3000/400 and between and 60% and 85% for the SPARCStation 20/51 we measured, relative to SPECInt92. Extensive cache simulations of three of these programs showed miss rates between less than 1% and 20% for different programs. However, based on a few microbenchmarks, we show that it is not cache behavior that causes degraded performance, but load delay cycles (on the Alpha), and inferior branch performance and software multiplication (on the Sparc).




Download Full History