Chapter 15: Analysis of Computer Networks Components

 < Free Open Study > 



Earlier chapters introduced the basic concepts and theories embodied in analytical modeling. Addressed were basic concepts in queuing systems theory, its application to computer systems modeling, and an introduction to network modeling. This chapter will address the use of analytical and simulation models specifically from the viewpoint of use as performance evaluation tools.

15.1 Introduction

In the past several years, the use of analytical performance models instead of the more widely used and familiar methods has become increasingly popular because of their relative simplicity of implementation and robustness of applications. These analytical models have been successful in estimation of such performance measures as throughputs, average queue lengths, and mean response times for a real system. This chapter is an introduction to queuing techniques for the modeling of computer communication networks, not an in-depth study.

The use of modeling to describe and imitate a real system has been with us since the beginning of the information revolution. These models are used not only to measure the performance of existing systems but also as part of the design and development of new systems. This latter goal is best attained through the use of analytical queuing models, as we will see in the following discussion of methods of performance evaluation.

The major performance evaluation tools (see Figure 15.1) other than queuing models are rules of thumb, linear projection, simulation, and benchmarking. These methods are listed in order of increasing complexity and implementation difficulty. The rules of thumb have been defined by the observation of operational systems and can be generally applied to local systems and extrapolated to distributed systems and networks. These rules take the following form:

click to expand
Figure 15.1: Spectrum of computer system modeling techniques.

  1. Generally, channel use in direct access storage devices (DASD) should not exceed 35 percent for on-line and 40 percent for batch applications.

  2. Individual DASD devices used should not exceed 35 percent.

  3. Average arm seek time on a DASD device should not exceed 50 cylinders.

  4. No block size for auxiliary storage should exceed 4 Kbytes.

These rules are useful in that they are easy to apply, economical to use, and can be applied to day-to-day operations. They are limited in the sense that they cannot be used to predict the usefulness of hardware or software upgrades.

The linear projection method has been used to pick up on the rules of thumb at the prediction limitation point. Although results can be obtained, the accuracy of the results is limited by the fact that a linear projection is used to predict the behavior of inherently nonlinear systems. This method also requires the availability of an existing system to measure the pertinent performance criteria to be used as a base for the projection and estimation of future resource requirements.

For simulation and benchmarking, there is no absolute distinction between development and implementation costs. Simulation allows the model to contain much more detail than the other methods, but this may not be an advantage when compared with queuing methods, where it has been found that too much information just serves to cloud the issue. Some simulation models are as large and cumbersome as the system they are modeling. The benchmarking method is the oldest and most used, but it is usually only helpful in the selection of the best hardware to process a known load. This is to say that the method requires existing hardware and, therefore, is not useful in the evaluation of hardware updates.

With the previous comments on other existing performance evaluation tools, we can assess the placement of queuing models and their overall usefulness. Queuing models reside between linear projection and simulation in terms of cost and complexity of implementation. Queuing models may be much simpler than the system they are modeling, because only the most pertinent performance parameters need to be accounted for. Not only do queuing models have a place in the evaluation of existing systems, but they also may be used in the design and development phase of new systems to help in the selection of hardware and hardware-software interaction to avoid system bottlenecks.

Recent advances in analytical modeling techniques are making analytical models increasingly capable of representing more and more aspects of the modeled system. Consequently, these techniques have been growing in popularity.

One method commonly used in system design is queuing analysis. Queuing models are more precise than other analytical techniques that predict performance based on average values [21]. One reason is that queuing models allow greater detail to be used in describing systems, and, hence, they capture the more important features of the system. Often, several submodels are required, as follows:

  1. Workload model. Specifies the characteristics of the resource demands on various equipment found in the system.

  2. Configuration or system structure model. Specifies the hardware characteristics of the system.

  3. Scheduling model. Specifies the scheduling algorithms whereby resources are allocated.

Queuing models can be categorized as either deterministic or stochastic in nature. If the design parameters to the model are known from prior experience or measurements, a deterministic analysis of the system may be carried out. Conversely, if the design parameters are not known, a stochastic analysis using various probability distributions is normally required.

Typical design parameters would include such items as:

  1. Interarrival rate of events

  2. Service times of these events

  3. Number of servers being modeled

  4. System capacity (i.e., number of events currently being processed and in queues)

  5. Queuing discipline employed (i.e., FIFO, LIFO, etc.)

Normally, queuing models provide some of the following performance attributes:

  1. Average queue lengths

  2. Average waiting time in queues

  3. Use statistics

  4. Average response times

Although queuing models have one overriding advantage in that they are cheap to use, there are a number of significant limitations to this method, as follows:

  1. Because these models assume the system has reached a steady state or equilibrium, peak or transient conditions are not modeled.

  2. These models are limited as to the complexity of the problems that can be solved. As problems become more complex or additional details are required, other methods must be used to model the systems.

  3. Without actually measuring various design parameters, it is difficult to determine whether the characteristics of the data used will represent the system under investigation.



 < Free Open Study > 



Computer Systems Performance Evaluation and Prediction
Computer Systems Performance Evaluation and Prediction
ISBN: 1555582605
EAN: 2147483647
Year: 2002
Pages: 136

flylib.com © 2008-2017.
If you may any questions please contact us: flylib@qtcs.net