An Empirical Study of Task Switching Locality in MVS

The "hit ratio" of a high-speed buffer (cache) depends on the "locality" of memory references. However, locality of reference is disturbed and the hit ratio decreases whenever a task switch occurs. This performance degradation can be minimized if "locality of task switching," the tendency for a small set of favored tasks to be frequently executed, exists and the cache is organized in such a way that it can hold blocks (lines) of multiple tasks. Locality of task switching and locality of memory references in individual tasks exhibit overall locality of memory references at a system level. This paper addresses the following questions. Does locality of task switching really exist? How can it be modeled? Task switching in IBM operating system/virtual storage with multiple virtual storage (OS/VS2 MVS) was measured using event traces for three different workloads to show that locality of task switching actually exists in MVS. Two different models of task switching are proposed. These models can be incorporated into cache multitasking models to predict more accurately the misses in real computer systems. A key parameter of these models is the task execution interval; measurements of execution intervals for the workloads used in the paper are presented and discussed.