TABLE OF CONTENTS Computer system design, in general, and memory hierarchy design, in particular, have reached a point at which it is no longer sufficient to design and optimize subsystems in isolation. A subsystem within a memory hierarchy could be a cache system, a DRAM system, or a disk system, but it could be much more fine-grained, such as a signaling scheme, a resource-management policy, a timing convention, a scheduling algorithm and any associated queue structures, a transaction protocol, a bus organization, a network topology, etc. Each of these items, even taken in isolation, is complex enough that it warrants significant research toward design optimization. When these subsystems are combined into a single, highly complex system, the resulting optimization problem is simply extraordinary.
Because memory systems are so complex, and because their subsystems are so complex, it is now the rule, and not the exception, that the subsystems we thought to be independent actually interact in unanticipated ways. These interactions frequently go undetected until late in the design cycle (e.g., at the time of implementation) and can cause significant cost and inconvenience or even errors of design that propagate into the installed product base.
To combat this problem, one must approach the design of even the smallest subsystem with an eye toward its relationship to the system as a whole and approach all system-level decisions with an understanding of all possible subsystem implications. One must make circuit-design decisions by considering their system-level impact. When designing interface protocols, one must quantify their system-level...
