Computer Architecture: a qualitative overview of Hennessy and Patterson

Computer Architecture is a wide-ranging subject, so it is useful to find a focus to make it interesting and to make sense of the detail. The modern approach to computer architecture research is to quantify as much as possible , so much of the material covered is about measurement, including how to report results and compare them. These notes however aim to provide a more qualitative overview of the subject, to place everything in context. The quantitative aspect is however important and is covered in the exercises. 1.1 Introduction The prescribed book contains a lot of detail. The aim of these notes is to provide a specific spin on the content as well as to provide a more abstract view of the subject, to help to make it easier to understand the detail. The main focus of this course is understanding the following two related issues: • the conflict between achieving latency and bandwidth goals • long-term problems caused by differences in learning curves This introductory chapter explains these concepts, and provides a starting point for understanding how the prescribed book will be used to support the course. First, the concepts are defined and explained. Next, this chapter goes on to break the course down into components, and then the components are related to contents of the book. Finally, the structure of the remainder of the notes is presented. This section looks at the main focus of the course, latency vs. bandwidth, and also introduces some of the major concepts in the course: what architecture is, the qualitative principle of design, and how performance is estimated or measured. 1.2.1 Latency vs. Bandwidth Loosely, latency is efficiency to the user, bandwidth is overall efficiency of the system. More accurately, latency is defined as time to complete a specific operation. Bandwidth (also called throughput) is the number of units of work that can be completed over a specific time unit. These two measures are very different. Completing a specific piece of work quickly to ensure minimum latency can be at the cost of overall efficiency, as measured by bandwidth. For example, a disk typically takes around 10ms to perform an access, most of which is time to move the head to the right place (seek time), and to wait for rotation of the disk (rotational delay). If a small amount of data is needed, the minimum latency is achieved if only that …