ITR: SafetyNet: Synergistic Support for Availability, Designability, Programmability, & Performance

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

EIA-0205286 David A. Wood University of Tennessee ITR: SafetyNet: Synergist Support for Availability, Designability, Programmability, & Performance Information services, supported by remote server computers, are becoming part of society's essential infrastructure for commerce, science, education, and government. Many servers are shared-memory multiprocessors that multiply the 60-fold single-processor performance improvements of the last decade by an additional factor of two to 64. These incredible performance improvements are putting increasing pressure on sever designs to improve complementary computer properties, such as: Availability. Society needs information services that are as dependable as our electricity and fresh water supplies. Unless new architectural techniques are developed and applied, however, server availability will go down with time, since deep submicron transistors are less reliable (and more susceptible to radiation) and future servers will employ more transistors, offering more failure opportunities. Designability. Designability is the challenge of deploying a completely correct computer. Getting the bugs out of today's computer systems is expensive, in both manpower and time, and getting worse as the transistor bonanza enables more complex designs. Furthermore, unpredictable delays leave expensive manufacturing facilities underutilized and cost 5% performance per month (due to Moore's Law). Programability. Servers typically run multi-threaded software to turn their raw computing power into high-throughput information services. Unfortunately, hard-to-test races make writing such software a challenge that stresses current programming practices and leads to bugs in deployed software. Hardware that facilitates rapid development of robust software should be prized. Performance. Solutions that improve availability, designability, and programmability will more likely flourish if they have, at most, a modest impact on cost-performance. Better yet would be hardware that actually improves performance, in addition to the other properties.