Scalable Shared-Memory Multiprocessing

9.^

Chapter Conclusions ^ 303

Overall, the results of the simulations are very positive. They show that if apphcations

contain enough parallehsm to exploit 2,048 processors, they wiU run weU on TeraDASH.

Even applications with high sharing rates (e.g., MP3D), which have relatively poor proces-

sor utilization in a single cluster, get a speedup of over 150 on 512 clusters. Furthermore, the

simulation shows that applications with reasonable miss rates do not cause excessive utihza-

tion of the cluster bus or the memory and directory banks. In particular, if each processor

executes for 37 active clocks between each miss (corresponding to 75 activ ...