Working sets and near-optimality

Abstract

In [1] Denning concludes from numerous observations of program behaviour that the WS algorithm with a single window size 8 is likely to "deliver throughput typically no worse than 10 percent from optimum". The authors of [2] report about observations of a set of programs, which requires several different window sizes for a i0 percent detuned WS algorithm. The purpose of this short note is to illustrate with the help of simple analysis why the observations reported in [1] were likely in the past and why observations like those of [2] become more and more likely in the future. To this end we introduce the notion of "degree of locality" and show that the locality set size and the degree of locality have impact on the optimal windowsize.