Tutorial

Abstract

The most important questions to be answered before attempting to monitor a machine are what to measure and why the measurement should be taken. There is no general answer to these questions, although a comprehensive set of considerations has been discussed elsewhere. The following example indicates some of the considerations involved. Suppose one is interested in tuning a medium scale system which utilizes virtual memory to support a batch multiprogramming strategy. The nature of the job load is a major factor in determining system performance; the mix may be monopolized by I/O-bound jobs which use very little processor time. In this case, the bottleneck might be the mass storage system or the peripheral devices. Resource utilization of the peripheral devices may indicate bottlenecks at that point; high mass storage utilization may not be attributable only to the I/O operations, but may be significantly influenced by the virtual memory replacement policy. Processor utilization in this system is also an insufficient measure for most purposes, since the overhead time for spooling, multiprogramming, and virtual memory may be unknown. A more useful measurement for operating system policy studies would quantify processor utilization for the user as well as for each function of interest in the operating system. From this example, one can see that the variety of evaluation objectives and computer systems causes the determination of what and why to be largely a heuristic problem.