One bottleneck encountered in large scale time-sharing systems (TSS) is excessive interactive swaps. The storage management of operating systems should minimize the number of physical interactive swaps. If working sets can be left resident until users complete their inputs, the number of swaps can be minimized. A demand swapping policy which maintains working sets in the main storage and swaps them only when a shortage of free storage space develops is a useful technique for resolving such bottlenecks. One important aspect of the demand swapping policy is the algorithm to determine which working set to swap out. To develop this demand swapping algorithm, the trace data of the behavior of actual TSS users at a terminal are accumulated and user input processes are analyzed. Five demand swapping algorithms (LRU, RAND, LUFO, PRED and SLRU) are proposed from the results. The number of physical interactive swaps that come from each demand swapping algorithm is compared using a trace-driven simulator. From this analysis, it is found that LRU is the best algorithms among the fixed-space demand swapping algorithms. However, SLRU, which is a kind of variable-space demand swapping algorithm, reduces the number of physical interactive swaps more than LRU within a given critical time range.
