Simulating working sets under MVS

Abstract

The concept of a "working-set" of a program running in a virtual memory environment is now so familiar that many of us fail to realize just how little we really know about what it is, what it means, and what can be done to make such knowledge actually useful. This follows, perhaps, from the abstract and apparently intangible facade that tends to obscure the meaning of working set. What we cannot measure of ten ranks high in curiosity value, but ranks low in pragmatic utility. Where we have measures, as in the page-seconds of SMF/MVS, the situation becomes even more curious: here a single number purports to tell us something about the working set of a program, and maybe something about the working sets of other concurrent programs, but not very much about either.This paper describes a case in which the concept of the elusive working set has been encountered in practice, has been intensively analyzed, and finally, has been confronted in its own realm. It has been trapped, wrapped, and, at last, forced to reveal it self for what it really is. It is not a number! Yet it can be measured. And what it is, together with its measures, turns out to be something not only high in curiosity value, but also something very useful as a means to predict the page faulting behavior of a program running in a relatively complex multiprogrmmed environment.The information presented here relates to experience gained during the conversion of a discrete event simulation model to a hybrid model which employs analytical techniques to forecast the duration of "steady-state" intervals between mix-change events in the simulation of a network-scheduled job stream processing on a 370/168-3AP under MVS. The specific "encounter" with the concept of working sets came about when an analytical treatment of program paging was incorporated into the model. As a result of considerable luck, ingenuity, and brute-force empiricism, the model won.Several examples of empirically derived characteristic working set functions, together with typical model results, are supported with a discussion of relevant modeling techniques and areas of application.