High-Performance IT Services

Terry Critchley

• Publisher: Auerbach Publications
• Year: 2016
• ISBN:
  9781315350103
• Edition: 1
• Topic: Business
• Subtopic: Operations
• Index: Business

This book on performance fundamentals covers UNIX, OpenVMS, Linux, Windows, and MVS. Most of the theory and systems design principles can be applied to other operating systems, as can some of the benchmarks. The book equips professionals with the ability to assess performance characteristics in unfamiliar environments. It is suitable for practitioners, especially those whose responsibilities include performance management, tuning, and capacity planning. IT managers with a technical outlook also benefit from the book as well as consultants and students in the world of systems for the first time in a professional capacity.

V APPENDICES

The appendices are here as backup material to the main body of this book. I have tried not to make the mistake of putting detail, above and beyond the call of understanding the topic in question at the desired level, in the body text. An example is the breakout of slightly more advanced (but necessary) queuing theory from the discussion in Chapter 7 and its placement in Appendix I. I became tired when reading books that drowned me in queuing theory before I could swim a few strokes in the subject. Most of it is purely academic anyway.

Similarly, I have included as an appendix a detailed discussion of some old benchmarks that may soon disappear from the literature on the subject to act as a record of things past.

Appendix I: Basic Queuing Theory

Introduction

Basic queuing theory stems to a large degree from the work of Erlang (ca. 1909) on telephony systems, using theory to predict and change the service perception of users of the systems. Queues and wait times are the factors that cause dissatisfaction in customers, which is bad for business.

It is therefore important to a business that the flow of work, queries, and other units of work or work units (WUs) pass through a system at an optimum rate, and queuing theory can help in this task. Some of the many factors of interest in queuing theory are listed below, not all of which are obvious to the customer but can have an effect on things he or she can witness, such as long wait times for service at a counter or supermarket checkout. Long, visible queues are bad for business.

A queue in a transaction processing system is not a visible problem unless it causes an unacceptable rise in response times. The upshot of this is that out of all the various quantities that can be derived via queuing theory, only some are of direct importance to any particular business or activity.

One thing to note about queuing is that, in general, there is a randomness about the traffic of work units arriving at a server for execution. These go under the name of stochastic processes. The other thing to note is that almost all quantities coming out of queuing calculations or simulations are averages (or means), and hence have a variation about this average or mean.

The queuing situations encountered can be classified thus:

■Constant arrivals, constant service times

■Random arrivals, constant service times

■Constant arrivals, random service times

■Random arrivals, random service times

The subject is made more difficult than simple math because the random arrivals and service profiles are described by math distributions that may or may not approximate to your own situations. In addition to this factor, there are others, such as queue limits—finite or infinite, multiple servers and multiple queues, and networks of queues.

Queuing Entities of Interest

■Type of system and environment

■Queuing discipline

■Population to be served

■Dispatching discipline within the servers

■Server service times

■Server idle time

■Number of servers

■Work unit (UoW)* arrival rate (λ)

■Service rate per server (μ)

■Service rate distribution (probability distribution function [pdf])

■WU arrival rate distribution (pdf)

■Average server utilization (ρ %), a measure of its busyness

■Peak server utilization (ρ_MAX %) before degradation occurs in either response time or throughput

■Peak-to-average utilization ratio (x:1), important in sizing exercises

■Average number of UoWs in the queue, for example, transactions

■Average number of UoWs in the system

■Average time a UoW spends in the queue (T_W) (waiting time)

■Average time a UoW spends in the server (T_S) (service time)

■Average total time for service (T_R), response time, which equals T_S+ T_W

■Probability that all servers are idle

■Probability that an arriving UoW waits

■Average rate of balking UoWs

■Maximum UoWs in the queue

■Average queue length

■Maximum queue length (capacity)

■Kendall’s notation

■Little’s theorem

■Transaction throughput rate

■Transaction response time and percentiles

■Batch work throughput

■Batch work job time

■Bottlenecks in workflow

Which entities are of interest depends on the organization involved. In transaction processing, the queue length isn’t a concern for the end user, only the resulting response time. However, for a supermarket planning checkouts and other facilities, a large, visible queue of customers creates a bad impression of the service offered and may affect whether they stay or even ever come again.

Service and Wait Times

In real life, total service times depend on the time spent waiting to be served (wait time) plus the time actually taken to be served (service time). These will depend on other factors, such as how busy the server is. In transactional systems, this time is often called the response time to a request. For example, if I enter a shop to buy a newspaper, the time I spend waiting to be served and the time taken to get and pay for the newspaper is called response time of the purchase of a newspaper transaction.

If the shop is empty and the newsagent is doing nothing else, there is no wait time and the system is said to be unconstrained; that is, there are no bottlenecks to impede service. Where there are bottlenecks resulting in wait times, the service is constrained. Thus, the equation for service time and wait time is

$$\begin{array}{ll}\text{Constrained service time}=\text{Unconstrined service time plus wait times}\hfill & (\mathrm{A1.1})\hfill \end{array}$$

This is the whole topic of latency, response time, wait times, and the rest in a nutshell.

Response Time: Unconstrained System

In an unconstrained system where utilizations are low ...

Table of contents

• Cover
• Half title
• Title Page
• Copyright Page
• Dedication
• Contents
• List of Figures
• List of Tables
• Foreword
• Preface
• Acknowledgments
• Author
• SECTION I PERFORMANCE PRIN CIPLES AND COMPONENTS
• SECTION II THE PERFORMANCE PLAYING FIELD
• SECTION III PERFORMANCE IN PRACTICE
• SECTION IV BENCHMARKS AND BENCHMARKING
• SECTION V APPENDICES
• Index