Ganger

Proceedings of the Hawaii International Conference On System Sciences, 1992, pp. 48-57

Robert Y. Hou, Gregory R. Ganger, Yale N. Patt
Department of Electrical Engineering and Computer Science, University of Michigan 

Charles E. Gimarc
NCR Corporation, E&M - Columbia 

It is well known that for many years, processor cycle times have continued to increase at a very rapid rate. On top of this, advances in multiprocessor technology have allowed potential system performance to increase at an even faster rate. The result, if we ignore the challenges in such fields as parallel algorithms, is that the performance of many of today's computer systems is limited by the I/O subsystem. In this paper, we attempt to do two things: 1) separate I/O space into three categories, based on their very different raisons d'etre and consequently very different characteristics, and 2) focus (the bulk of the paper) on the issues pertaining to improving the performance of one basic mechanism in the I/O subsystem, the magnetic disk. We do the former in order to set the framework of the I/O space. We believe that if we are to improve the performance of the I/O subsystem, we first have to understand the nature of I/O and not cloud our effeorts by treating I/O as one homogeneous structure. We do the latter as the first step in dealing with the various mechanisms that make up the I/O space.

Download: pdf