The NUCAR Army Publications General Information Conferences Research Projects Home

Memory Hierarchy Performance Research Students: Recent Participants: Jason Casmira Mona Dimitri (Raytheon) Yue Liu Cascade (Communications) Faculty: David Kaeli Project Summary: C ache memories exploit the properties of temporal and spatial locality exhibited by applications developed in high level languages. The instruction stream exhibits high temporal and spatial locality patterns, while the data stream does not. In this research, we investigate prefetching techniques for the the data stream. We begin by inspecting the high-level language data structures which commonly cause problems in data caches (e.g., arrays, linked lists). We also have developed methods to avoid the cache pollution introduced by errant prefetching. We propose a new mechanism called the Prefetch Buffer Filter, which intercepts prefetched blocks before commiting them to the data cache. The PBF has been shown to be highly effective in reducing the amount of data cache pollution. Publications and further information: " Modeling Cache Pollution ", Appearing in the Proceedings of IASTED International Conference on Modeling and Simulation, 1995 Jason P. Casmira and David R. Kaeli " Branch-directed and Pointer-based Data Cache Prefetching ", To appear in the Journal of Systems Architecture 1998 Yue Liu, Mona Dimitri, and David R. Kaeli