Some experiments in simulating OS/360 from performance data

Abstract

Recent work has indicated that the time a program spends in memory is the primary determinant of the turnaround of the program.We constructed a simulation model of our 0S/360 system, using as input performance data gathered by a software monitor, and experimented with differing models of the operating system, program behavior, and hardware configurations to determine an appropriate trade-off between the complexity of models used and the accuracy of the simulation.We constructed a fairly simple model which simulated, from resource requirements of programs, the time a collection of programs resided in memory to within 10% accuracy for the actual completion time of the various programs.The data gathered gave, between any two sample points, the central processor time spent processing on interrupts, on various system tasks, on various problem program tasks and on the large timesharing task which runs in our system. It also gave the number and total duration of the input-output requests from each task to each device. This data was sampled every time a program was introduced into memory or terminated and left memory.The simulators we constructed had components for the operating system (in various levels of detail), problem program behavior (we tested various distributions of processor burst time and input-output time) and also a component for the inputoutput configuration (we tested various models of channels and devices with associated queueing).The test of the accuracy of the simulation Was how well it compared with the times taken by the actual programs.The results yielded insight into the nature of program behavior as well as how programs interact with the operating system.