Techniques for improving the perfomance and scalability of directory-based shared-memory multiprocessors

Abstract

Cache-coherent, nonumiform memory acces or cc-NUMA is an attractive architecture for building a spectrum of shared memory multiprocessors (whic are socing widespread use in commercial, technical and scientific applications). Unfortunately, there are some factors which limit the maximun number of processors that can be offered at a good price/perfomance ratio. This paper presents a survey of some of the proposals that have recently appeared focusing on tow of these factors: the increased cost, in terms of hardware overhead that the use of directories entails, and the long cache miss latencies obseved in these designs as a consequence of the indirection introduced by the acces to the directory.