This paper appears in the March, 1972, issue of the Communications of the ACM. Its abstract is reproduced below.Five well-known scheduling policies for movable head disks are compared using the performance criteria of expected seek time (system oriented) and expected waiting time (individual I/O request oriented). Both analytical and simulation results are obtained. The variance of waiting time is introduced as another meaningful measure of performance, showing possible discrimination against individual requests. Then the choice of a utility function to measure total performance including system oriented and individual request oriented measures is described. Such a function allows one to differentiate among the scheduling policies over a wide range of input loading conditions. The selection and implementation of a maximum performance two-policy algorithm are discussed.(Pages 115 through 121 omitted) 114

Five well-known scheduling policies for movable head disks are compared using the performance criteria of expected seek time (system oriented) and expected waiting time (individual I/O request oriented). Both analytical and simulation results are obtained. The variance of waiting time is introduced as another meaningful measure of performance, showing possible discrimination against individual requests. Then the choice of a utility function to measure total performance including system oriented and individual request oriented measures is described. Such a function allows one to differentiate among the scheduling policies over a wide range of input loading conditions. The selection and implementation of a maximum performance two-policy algorithm are discussed.

Five well-known scheduling policies for movable head disks are compared using the performance criteria of expected seek time (system oriented) and expected waiting time (individual I O request oriented). Both analytical and simulation results are obtained. The variance of waiting time is introduced as another meaningful measure of performance, showing possible discrimination against individual requests. Then the choice of a utility function to measure total performance including system oriented and individual request oriented measures is described. Such a function allows one to differentiate among the scheduling policies over a wide range of input loading conditions. The selection and implementation of a maximum performance two-policy algorithm are discussed.

This paper presents a method for dynamic modification of operating system control parameters to improve system performance. Improved parameter settings are learned by experimenting on the system. The experiments compare the performance of alternative parameter settings in each region of a partitioned load-performance space associated with the system. The results are used to modify important control parameters periodically, responding to fluctuations in system load and performance. The method can be used to implement adaptive tuning, to choose between alternative algorithms and policies, or to select the best fixed settings for parameters which are not modified. The method was validated and proved practical by an investigation of two parameters governing core quantum allocation on a Sperry Univac 1100 system. This experiment yielded significant results, which are presented and discussed. Directions for future research include automating the method, determining the effect of simultaneous modifications to unrelated control parameters, and detecting dominant control parameters.

A software measurement facility which is part of a general purpose time-sharing system is described. The Data Collection Facility (DCF) has been implemented in the Michigan Terminal System (MTS) for the System 360 model 67. It exists for the purpose of monitoring operating system and user program behavior and performance. The overall structure of MTS is outlined in order to explain the implementation of the DCF. Events in the system are identified and recorded from within the supervisor, and dumped to magnetic tape by an auxiliary program for off-line processing. Events in user programs which are unrelated to system actions are recorded with a supervisor call. The time of occurrence of each event is accurately recorded, and data items are furfher identified by job and type. The overhead associated with data collection and its interference with normal jobs is carefully analyzed, and both are shown to be minimal. Several examples are given of information obtained with the facility and of applications in which it has been useful. Some general guidelines are offered for the construction of future monitoring programs. IntroduclioiiThe intent of this paper is to describe the structure, operation, and experience with a software data eollection facility (DCF) which was written for a time-shared operating system to provide behavior and performance data on the operation of the sj'stem and u.ser programs. The DCF is a mechanism for the collection of a selected series of event descriptions identified by time, job, and type of event, and consisting of information relevant to the event t^T^e. These data items are collected in buffers and dumped onto magnetic tape for subsequent "off-line" processing, which includes separation of events by job, interpretation of common job scheduling events to define job status, and isolation of job service requests from system actions.Event recording in operating systems is certainly nothing new. The trace table, the use of .system cells to count occurrences of events, and records kept for timeaccounting are examples of the technique; they are all used in the Michigan Terminal System (MTS) as well aa the DCF. It is widely believed, however, that a general purpose facility cannot be provided for a large scale multiprogramming system without squandering valuable resources and thus compromising the inforniation obtained. This is not so.Although the DCF was written to provide specific data for the author's own research, the implementation has not restricted the type of events which can be recorded. Among other things, it has been used to determine operating system overhead faetors, measure individual resource utilization, verify hardware performance, and measure its own interference. Events with a few microseconds duration up to those of several hours in length have been clocked, and the resulting information has proved useful for system "fine tuning" and long-range system design.Two software measurement techniques which have recently been reported [1, 2] are based on sampling rather than continuo...