Overview

Memory is essential to the operation of a computer system, and nothing is more important to the development of the modern memory system than the concept of the memory hierarchy. While a flat memory system built of a single technology is attractive for its simplicity, a well-implemented hierarchy allows a memory system to approach simultaneously the performance of the fastest component, the cost per bit of the cheapest component, and the energy consumption of the most energy-efficient component.

For years, the use of a memory hierarchy has been very convenient, in that it has simplified the process of designing memory systems. The use of a hierarchy allowed designers to treat system design as a modularized process to treat the memory system as an abstraction and to optimize individual subsystems (caches, DRAMs [dynamic RAM], disks) in isolation.

However, we are finding that treating the hierarchy in this way as a set of disparate subsystems that interact only through well-defined functional interfaces and that can be optimized in isolation no longer suffices for the design of modern memory systems. One trend becoming apparent is that many of the underlying implementation issues are becoming significant. These include the physics of device and interconnect scaling, the choice of signaling protocols and topologies to ensure signal integrity, design parameters such as granularity of access and support for concurrency, and communication-related issues such as scheduling algorithms and queueing. These low-level details have begun to affect the higher level design process quite dramatically, whereas they were considered transparent only...