Experimental results on the paging behavior of numerical programs

Traces of numerical programs are used to examine their behavior in a paged virtual memory system. The working set policy is used for the replacement algorithm. It is found that the behavior of such programs is different from the behavior of other types of programs like compilers and system programs. These differences are most significant in the lifetime curves and the space time cost curves. All programs examined showed ill-behavior. Moreover, the space-time costs of executing these programs are very sensitive to the choice of the control parmater, the window size. Our measurements show that approximations based on the common practice of using virtual time instead of real time in generating statistics are often inaccurate. The "primary knee criterion" of optimizing the space-time cost did not hold for some programs. The parameter-real memory and the real memory-fault rate anomalies show significantly in all but one of the seventeen programs examined. Index terms: virtual memory, the working set policy, numerical program behavior, lifetime curves, space-time cost curves, optimal multiprogramming, the primary knee criterion, sensitivity to control parameter, virtual time vs. real time analysis, working set anomalies.