Develop your computer performance pattern
Jerre D. Noe, N. W. Runstein · ACM SIGMETRICS Performance Evaluation Review · 1974
Is the load on your computer shifting? Did that change to faster access disks really help? Would more core memory increase throughput appreciably, or would it be necessary to also increase central processor power? These are three quite different kinds of questions; one concerns detecting a long-term trend, another assessing the effects of a system change, and a third estimating effects of the decision to alter the configuration. Yet all of these require knowledge of current and past system performance, the type of knowledge that must be the result of long-term performance monitoring. This is not simple enough to be picked up overnight or in one series of experiments, nor can it be assessed by watching one or two parameters over a long period. One must have a thorough understanding of the pattern of performance by knowing the mean values of a number of measures and knowing something about the variations from these means. This paper hardly needs to recommend that computer managers establish an understanding of performance pattern; they already are very conscious of the need. What it does is recount development of a method of doing so for the CDC 6400 at the University of Washington and of the selection of “Kiviat Graphs” as a means to present data in a synoptic form. The remainder of this paper will give a brief account of the authors' experience in designing a measurement system for the CDC 6400 at the University of Washington Computer Center. This will include comments on the approach to deciding what to measure an d display for the synoptic view of the system, as well as how to provide more detailed data for backup. Examples of the use of Kiviat Graphs [4] to show the effects of load shift and of a system configuration change are included, and the effect of a change of operating system will be noted.