A Study of the Memory Wall within the Jacobi Iteration Method
Siqi Sun, Shan Wang, Wenfeng Shen, Weimin Xu, Yanheng Zheng · 2012
In recent years, a great number of applications have been implemented on the CMP and achieved good performance. The success of the parallelism of a great deal of applications on CMP shows a bright future of the development of the CMP. However, some other experiments exposed the shortcomings of the CMP which may significantly reduce the performance of the parallel program. In this paper, we devote the effort to reveal the effect of the memory wall on the performance on different types of CMP through a data-intensive application, the Jacobi iteration method. The memory wall is becoming a more and more crucial issue that dominates the performance of the whole application, especially data-intensive ones. High latency, low bandwidth and data contention compose the memory wall in the CMP. Limiting the problem scale within a certain range to avoid the insufficient bandwidth is a feasible way. Some advanced technologies such as data prefetching and thread switching can help reducing or even eliminating the effect of the latency.