New directions in workload characterization
Thomas D. Wagner · 1993
Workload characterization is the process of building models of computer system workloads. The problems of workload characterization received much attention over a decade ago. With the development of parallel systems this subject requires re-examination. New methods for workload characterization are developed. Methods for use with traditional systems as well as methods for use with parallel multiprocessor systems are examined. The existing characterizations of parallel workloads are compared and contrasted and mappings between them are developed. A method for determining the workload demand parameters that describe a multi-class queueing network is presented. This new exponential sieve technique is shown to perform better than traditionally accepted techniques. The workload model that is constructed using this technique is used as input to queueing models and is shown useful in constructing performance prediction models. A workload characterization based on segregation measures is introduced. This characterization expresses the error that can be made when modelling multi-class systems with single class queueing models. The behavior of this segregation measure is shown to depend on the variance in mean demand in the workload. This variance in mean demand is a measure of system balance. The segregation measure is also shown to depend on a new measure that expresses how close measured demand data is to single class. The evolution and widespread use of MIMD multiprocessors is accompanied by new challenges for system modellers. This, in turn, implies new challenges in workload modelling. The prediction of speedup, as additional processors are assigned to a parallel workload, is one such challenge. A new algorithm that identifies phases of homogeneous processor utilization in the execution profile of parallel workloads is presented. By identifying these phases, predictions of the speedup characteristics of a parallel program are made more accurately. Another challenge of MIMD multiprocessor systems is the processor allocation problem. Stochastic learning automata are applied to the processor allocation problem. In this use of stochastic learning automata, the model of the system and the model of the workload must be combined. Through simulation, this new approach is shown to be an effective solution to the problem of how many processors, from a pool of parallel processors, should be allocated to an arriving job so that total system power is maximized.