Improving performance and energy consumption in region-based caching architectures.

Abstract

Embedded systems must simultaneously deliver high performance and low energy consumption. Meeting these goals requires customized designs that fit the requirements of the targeted applications. This philosophy of tailoring the implementation to the domain applies to all subsystems in the embedded architecture. For the memory system, which is a key performance bottleneck and a significant source of energy consumption, generic caching strategies are insufficient. The system requires specialized cache structures that match the manner in which programmers use data. Since different data subsets exhibit varying degrees of locality, a partitioned cache offers the best opportunity to optimize performance and energy consumption for all memory accesses. In this dissertation, I explore several different methods that utilize partitioning for an energy-efficient, high performance data cache. Region-based caching, which replaces a unified data cache with multiple caches optimized for stack, global, and heap references, serves as the starting point for this research. I begin by addressing energy consumption in the level one data cache. Drowsy region-based caches combine static and dynamic energy reduction techniques to simultaneously lower both sources of energy consumption. This combination performs better than the sum of its parts because the separate region caches allow us to use different degrees of drowsy caching. I then show how additional cache partitioning can further reduce energy consumption, presenting a scheme to identify highly local data in the heap region and route their accesses to a smaller cache. I also study methods for improving memory system performance. The effectiveness of data prefetching can be increased by partitioning the cache in a manner that isolates data that prefetch well. Finally, I discuss how to reallocate data within region-based caches to eliminate unnecessary cache misses.