Proceedings of the CMG'89 Conference
Experiences of Hierarchical Workload Modeling in Capacity Planning
Kimmo E. E. Raatikaineny
University of Helsinki, Department of Computer Science Teollisuuskatu 23, SF-00510 Helsinki, Finland
In an extensive research project we have developed and analyzed modeling tools and techniques to be
used in practical situations of capacity planning. Our objective is to construct queueing network models
that can be used to analyze the performance of a VAXCluster-system both with measured and with
The reported experiences indicate that the hierarchical modeling of workload provides a fruitful basis for integrating the capacity planning into organization-wide general planning. However, the standard measurement tools of VAX/VMS are insufficient to identify all parameters of the queueing network workload. Considerable specialized knowledge of the examined system, carefully planned and executed benchmarks, and subjective adjustments are required in identifying the io-workload.
In principle, a capacity planning study consists of three primary steps. Current and future business activities identify the needs and opportunities for employing computerized information processing. These needs and opportunities are the external premises of capacity planning. The first step of a capacity planning study is to translate these premises into computing requirements and performance objectives. When the requirements and objectives are specified, the second step is to predict and analyze the performance of possible computer system configurations. Finally, the third step utilizes performance predictions and cost-benefit analysis. They are the basis of constructing the decision alternatives for the development of the information processing. In this paper we consider the computing requirements or more precisely how activities of an organization can be translated into workload of a queueing network model.
We regard capacity planning as an integral part of general organization-wide planning. This implies that business activities of the organization must be translated into the consumption of various hardware resources. Since business activities have no direct correspondence to the consumption of hardware resources, we use a hierarchical model of the workload. Business activities are first translated into a functional workload . The characterization of workload as service demands of hardware resources is necessary in performance predictions that are based on queueing network analysis.
The characterization of workload involves three phases. Usually the workload is so heterogeneous that the queue-
y This work was supported by Emil Aaltonen Foundation
ing network model must have multiple classes of customers. This classification is the first phase. The other two phases consist of identifying the functional and queueing network workload .
The second practical factor that has affected our modeling and identification is that the monitoring of the current workload and performance should not create additional load . Therefore we have decided to use primarily the data that is automatically collected. In the examined system measurement data is provided by the VAX/VMS Accounting and Monitor Utilities.
In this paper we begin by discussing modeling issues. We review briefly some specific features of the examined system and queueing network modeling. In Section 3, we summarize the measurement data obtained by VAX/VMS Accounting and Monitor Utilies. In Section 4 we concentrate on workload modeling. In this paper we consider all three phases of workload characterization: functional classification, identification of functional workload, and identification of queueing network workload.
2. Modeling Issues
In modeling the basic concern is the system under consideration. The system's specific features; the hardware, workload, and measurement tools; determine to a great extent possibile methods used for modeling. In this section we summarize the specific features of the examined system and discuss queueing network modeling.
Our approach to modeling is based on two fundamental principles. The first one is decomposition of the performance modeling effort. The second one is hierarchical