Improving energy and performance of data cache architectures by exploiting memory reference characteristics.
Hsien-Hsin Sean Lee, Gary Tyson · Deep Blue (University of Michigan) · 2001
Minimizing power, increasing performance, and delivering effective memory bandwidth are today's primary microprocessor design goals for the embedded, high-end and multimedia workstation markets. In this dissertation, I will discuss three major data cache architecture design optimization techniques, each of which exploits the data memory reference characteristics of the applications written in high-level languages. Through a better understanding of the memory reference behavior, we can design a system that executes at higher performance, while consuming less energy, and delivering more effective memory bandwidth. The first part of this dissertation presents an in-depth characterization of data memory references, including analysis of semantic region accesses and behavior of data stores. This analysis leads to a new organization of the data cache hierarchy called Region-based Cachelets. Region-based Cachelets are capable of improving memory performance of embedded applications while significantly reducing dynamic energy consumption, resulting in a 50% to 70% improvement in energy-delay product efficiency using this approach. Following this, I will discuss a new cache-like structure, the Stack Value File (or SVF), which boosts performance of general purpose applications by routing stack data references to a separate storage structure optimized for the unique characteristics of the stack reference substream. By utilizing a custom structure for stack references, we are able to increase memory level parallelism, reduce memory latency, and reduce off-chip memory activity. The performance can be improved by 24% by implementing an 8KB SVF for a processor with a dual-ported L1 cache. Finally, I will address memory bandwidth issues by proposing a new write policy called Eager Writeback which can effectively improve overall system performance by shifting the writings of dirty cache lines from on-demand to times when the memory bus is less congested. It lessens the criticality of on-demand misses and improves performance by 6% to 16% for the 3D graphics geometry pipeline.