Performance Requirements: CMG'11 slides with notes (pdf)

Performance requirements are supposed to be tracked from the system inception
through the whole system lifecycle including design, development, testing,
1
through the whole system lifecycle including design, development, testing,
operations, and maintenance. However different groups of people are involved on
each stage using their own vision, terminology, metrics, and tools that makes the
subject confusing when going into details. The presentation is an attempt of a
systematic view of the subject.

Performance requirements are supposed to be tracked from the system
inception through the whole system lifecycle including design, development,
2
inception through the whole system lifecycle including design, development,
testing, operations, and maintenance. However different groups of people are
involved on each stage using their own vision, terminology, metrics, and tools that
makes the subject confusing when going into details.
For instance, business analysts use business terms. The architects’
community uses its own languages and tools (mostly created for documenting
functionality so performance doesn’t fit them well). Developers often think about
performance through the profiler view. The virtual user notion is central for
performance testers. Capacity planners use some mathematical terminology when
they come up with queuing models. Production people have their own tools and
metrics; and executives are more interested in high-level, aggregated metrics. These
views are looking into the same subject – system performance – but through
different lenses and quite often these views are not synchronized and differ
noticeably. All of these views should be synchronized to allow tracing performance
through all lifecycle stages and easy information exchange between stakeholders.
Many existing approaches to describing performance requirements try to put these
multi-dimensional and cross-dependent performance views into a set of simple flat
templates designed for functional requirements.
The presentation provides an overview of current issues and is an attempt to
create a holistic view of the subject.
Disclaimer: The views expressed here are my personal views only and do not necessarily represent those of my current or
previous employers. All brands and trademarks mentioned are the property of their owners.

First we discuss different performance metrics and terms, and then we look
through the performance requirements process following IEEE Softwarethrough the performance requirements process following IEEE Software
Engineering Book of Knowledge (SWEBOK) terminology [SWEBOK04].
3

Let’s take a high-level view of a system. On one side we have users who use
the system to satisfy their needs. On another side we have the system, a combinationthe system to satisfy their needs. On another side we have the system, a combination
of hardware and software, created (or to be created) to satisfy user’s needs.
Users are not interested in what is inside the system and how it functions as
soon as their requests get processed in a timely manner (leaving aside personal
curiosity and subjective opinions).
4

So business requirements should state how many requests of each kind go
through the system (throughput) and how quickly they need to be processedthrough the system (throughput) and how quickly they need to be processed
(response times). Both parts are vital: good throughput with long response times
usually is as unacceptable, as are good response times with low throughput. While
throughput is definitely comes from business, response times are rather usability
requirements when they are good enough, but starting to impact business as soon as
they become not so good.
5

Throughput is the rate at which incoming requests are completed.
Throughput defines the load on the system and is measured in operations per time
6
Throughput defines the load on the system and is measured in operations per time
period. It may be the number of transactions per second or the number of processed
orders per hour. In most cases we are interested in a steady mode when the number
of incoming requests would be equal to the number of processed requests.
Defining throughput may be pretty straightforward for a system doing the
same type of business operations all the time, like processing orders or printing
reports when they are homogenous. Clustering requests into a few groups may be
needed if requests differ significantly: for example, small, medium, and large
reports. It may be more difficult for systems with complex workloads; the ratio of
different types of requests can change with the time and season.
Throughput usually varies with time. For example, throughput can be
defined for a typical hour, peak hour, and non-peak hour for each particular kind of
load. In environments with fixed hardware configuration the system should be able
to handle peak load, but in virtualized or cloud environments it may be helpful to
further detail what the load is hour-by-hour to ensure better hardware utilization.

Quite often, however, the load on the system is characterized by the number
of users. Partially it is coming from the business (in many cases the number of usersof users. Partially it is coming from the business (in many cases the number of users
is easier to find out), partially it is coming from performance tests: unfortunately,
quite often performance requirements get defined during performance testing and
the number of users is the main lever to manage load in load generation tools.
But the number of users doesn’t, by itself, define throughput. Without
defining what each user is doing and how intensely (i.e. throughput for one user),
the number of users doesn’t make much sense as a measure of load. For example, if
500 users are each running one short query each minute, we have throughput of
30,000 queries per hour. If the same 500 users are running the same queries, but
only one query per hour, the throughput is 500 queries per hour. So there may be the
same 500 users, but a 60X difference between loads (and at least the same
difference in hardware requirements for the application – probably more,
considering that not all systems achieve linear scalability).
7

Homogenous throughput with randomly arriving requests (sometimes assumed in
modeling and requirements analysis) is a simplification in most cases. In addition tomodeling and requirements analysis) is a simplification in most cases. In addition to
different kind of requests, most systems use a kind of sessions: some system resources are
associated with the user (source of requests). So the number of parallel users (sessions)
would be an important requirement further qualifying throughput. In a more generic way
this metric may be named concurrency: the number of simultaneous users or threads. It is
important: connected but inactive users still hold some resources.
The number of online users (the number of parallel sessions) looks like the best
metric for concurrency (complementing throughput and response time requirements).
However terminology is somewhat vague here, sometimes “the number of users” may have
a completely different meaning:
• Total or named users (all registered or potential users). This is a metric of data the
system works with. It also indicates the upper potential limit of concurrency. In some cases
may be used as a way to find out concurrency as a percentage of total user population, but
definitely is not a concurrency metric.
• “Really concurrent” users: the number of users running requests at the same time.
In most cases it is matching the number of requests in the system. While that metric looks
appealing, it is not a load metric: the number of "really concurrent" requests depends on the
processing time for this request. The shorter is processing time, the fewer concurrent
requests we have in the system. For example, let’s assume that we got a requirement to
support up to 20 “concurrent” users. If one request takes 10 sec, 20 “concurrent” requests
mean throughput of 120 requests per minute. But here we get an absurd situation that if we
improve processing time from 10 to one second and keep the same throughput, we miss our
requirement because we have only two “concurrent” users. To support 20 “concurrent”
users with a one-second response time, we really need to increase throughput 10 times to
1,200 requests per minute.
It is important to understand what users we are discussing: the difference between
each of these three “number of users” metrics may be drastic.
8

Response times (in the case of interactive work) or processing times (in the
case of batch jobs or scheduled activities) define how fast requests should becase of batch jobs or scheduled activities) define how fast requests should be
processed. Acceptable response times should be defined in each particular case. A
time of 30 minutes could be excellent for a big batch job, but absolutely
unacceptable for accessing a Web page in a customer portal. Response times depend
on workload, so it is necessary to define conditions under which specific response
times should be achieved; for example, a single user, average load or peak load.
Response time is the time in the system (the sum of queuing and processing
time). Usually there is always some queuing time: just because the server is a
complex object with sophisticated collaboration multiple components including
processor, memory, disk system, and other connecting parts. That means that
response time is larger than service time (to use in modeling) in most cases.
Significant research has been done to define what the response time should
be for interactive systems, mainly from two points of view: what response time is
necessary to achieve optimal user’s performance (for tasks like entering orders) and
what response time is necessary to avoid Web site abandonment (for the Internet).
Most researchers agreed that for most interactive applications there is no point in
making the response time faster than one to two seconds, and it is helpful to provide
an indicator (like a progress bar) if it takes more than eight to 10 seconds.
Response times for each individual transaction vary, so we need to use some
aggregate values when specifying performance requirements, such as averages or
percentiles (for example, 90 percent of response times are less than X). Apdex
standard uses a single number to measure user satisfaction [Apdex].
For batch jobs, it is important to specify all schedule-related information,
including frequency (how often the job will be run), time window, dependency on
other jobs and dependent jobs (and their respective time windows to see how
changes in one job may impact others).
9

It is very difficult to consider performance (and, therefore, performance
requirements) without full context. It depends, for example, on the volume of data
10
requirements) without full context. It depends, for example, on the volume of data
involved, hardware resources provided, and functionality included in the system. So
if any of that information is known, it should be specified in the requirements. Not
everything may be specified at the same point: while the volume of data is usually
determined by the business and should be documented at the beginning, the
hardware configuration is usually determined during the design stage.

The performance metrics of the system (the right side of the high-level
view) are not important from the business (or user) point of view, but are very
11
view) are not important from the business (or user) point of view, but are very
important for IT (people who create and operate the system). These internal
(technological) requirements are derived from business and usability requirements
during design and development and are very important for the later stages of the
system lifecycle. Traditionally such metrics were mainly used for monitoring and
capacity management because they are easier to measure and only recently tools
measuring end-user performance get some traction.

The most wide-spread metric, especially in capacity management and
production monitoring, is resource utilization. The main groups of resources are
12
production monitoring, is resource utilization. The main groups of resources are
CPU, I/O, memory, and network. However, the available hardware resources are
usually a variable in the beginning. It is one of the goals of the design process to
specify hardware needed for the system from the business requirements and other
inputs like company policies, available expertise, and required interfaces.
When resource requirements are measured as resource utilization, they are
related to a particular hardware configuration. They are meaningful metrics when
the hardware configuration is known. But these metrics doesn’t make any sense as
requirements until the hardware configuration would be decided upon: how we can
talk, for example, about processor utilization if we don’t know yet how many
processors we would have? And such requirements are not useful as requirements
for software if it get deployed to different hardware configurations, and, especially,
for Commercial Off-the-Shelf (COTS) software.
Only way we can speak about resource utilization on early phases of the
system lifecycle is as a generic policy. For example, corporate policy may be that
CPU utilization should be below 70 percent.

When required resources are specified in absolute values, like the number of
instructions to execute or the number of I/O operations per transaction (asinstructions to execute or the number of I/O operations per transaction (as
sometimes used, for example, for modeling), it may be considered as a performance
metric of the software itself, without binding it to a particular hardware
configuration. In the mainframe world, MIPS was often used as such metric for
CPU consumption, but there is no such widely used metric in the distributed
systems world.
The importance of resource-related requirements is increasing again with
the trends of virtualization, cloud computing, and service-oriented architectures.
When we depart from the “server(s) per application” model, it becomes difficult to
specify requirements as resource utilization, as each application will add only
incrementally to resource utilization. There are attempts to introduce such metrics.
For example, the ‘CPU usage in MHz’ or ‘usagemhz’ metric used in the VMware
world or the ‘Megacycles’ metric sometimes used by Microsoft [Microsoft10].
Another related metric sometimes (but rarely) used is efficiency when it is defined
as throughput divided by resources (however the term is often used differently).
In the ideal case (for example, when the system is CPU bound and we can
scale the system linearly just adding processors) we can easily find hardware
configuration needed if we have an absolute metric of resources required.
For example, if software needs X units of hardware power per request and a
processor has Y units of hardware power, we can calculate the number of such
processors N needed for processing Z requests as N=Z*X/Y. The reality, of course,
is more sophisticated. First of all, we have different kinds of hardware resources:
processors, memory, I/O, and network. Usually we concentrate on the most critical
one keeping in mind others as restrictions.
13

Scalability is system's ability to meet the performance requirements as the
demand increases (usually by adding hardware). Scalability requirements may
14
demand increases (usually by adding hardware). Scalability requirements may
include demand projections such as an increasing of the number of users,
transaction volumes, data sizes, or adding new workloads. How response times will
increase with increasing load or data is important too (load or data sensitivity).
From a performance requirements perspective, scalability means that you
should specify performance requirements not only for one configuration point, but
as a function of load or data. For example, the requirement may be to support
throughput increase from five to 10 transactions per second over next two years
with response time degradation not more than 10 percent.
Scalability is also a technological (internal IT) requirement. Or perhaps even
a “best practice” of systems design. From the business point of view, it is not
important how the system is maintained to support growing demand. If we have
growth projections, probably we need to keep the future load in mind during the
system design and have a plan for adding hardware as needed.

IEEE Software Engineering Book of Knowledge defines four stages for
requirements [SWEBOK04]:requirements [SWEBOK04]:
• Elicitation: Where requirements come from and how to collect them.
• Analysis: Classify / Elaborate / Negotiate.
• Specification: Production of a document. While documenting requirements
is important, the way to do this depends on software development methodology
used, corporate standards, and other factors.
• Validation: making sure that requirements are correct.
16

If we look at the performance requirements from another point of view, we
can classify them into business, usability, and technological requirements.
17
can classify them into business, usability, and technological requirements.

Business requirements come directly from the business and may be captured
very early in the project lifecycle, before design starts. For example, "a customervery early in the project lifecycle, before design starts. For example, "a customer
representative should enter 20 requests per hour and the system should support up to
1000 customer representatives". If translated into more technical terms, requests
should be processed in five minutes on average, throughput would be up to 20,000
requests per hour, and there could be up to 1,000 parallel sessions.
The main trap here is to immediately link business requirements to a
specific design, technology, or usability requirements, thus limiting the number of
available design choices. If we consider a Web system, for example, it is probably
possible to squeeze all the information into a single page or have a sequence of two
dozen screens. All information can be saved at once or each page of these two-
dozen can be saved separately. We have the same business requirements, but
response times per page and the number of pages per hour would be different.
18

While the final requirements should be quantitative and measurable, it is not
an absolute requirement for initial requirements. Scott Barber, for example,an absolute requirement for initial requirements. Scott Barber, for example,
advocates that we need to gather qualitative requirements first [Barber07]. While
business people know what the system should do and may provide some numeric
information, they are not trained in requirement elaboration and system design. If
asked to provide quantitative and measurable requirements, they finally provide
them based on whatever assumptions they have about system's design and human-
computer interaction, but quite often this results in wrong assumptions being
documented as business requirements. We need to document real business
requirements in the form they are available, and only then elaborate them into
quantitative and measurable.
19

One often missed issue, as Scott Barber notes, is goals versus requirements
[Barber07]. Most of response time “requirements” (and sometimes other kinds of
20
[Barber07]. Most of response time “requirements” (and sometimes other kinds of
performance requirements,) are goals, not requirements: something that we want to
achieve, but missing them won’t necessarily prevent deploying the system.
In many cases, especially for response times, there is a big difference
between goals and requirements (the point when stakeholders agree that the system
can’t go into production with such performance). For many interactive web
applications, response time goals are two to five seconds and requirements may be
somewhere between eight seconds and one minute.
One approach may be to define both goals and requirements. The problem is
that, except when coming from legal or contractual obligation, requirements are
very difficult to get. Even if stakeholders define performance requirements, quite
often, when it comes to go/no go decisions, it becomes clear that it was not the real
requirements, but rather second-tier goals.

In addition, multiple performance metrics only together provide the full
picture. For example, you may have a 10-second requirement and you get 15-secondpicture. For example, you may have a 10-second requirement and you get 15-second
response time under the full load. But what if you know that this full load is the high
load on the busiest day of year, that response times for the maximal load for other
days are below 10 second, and you see that it is CPU-constrained and may be fixed
by a hardware upgrade? Real response time requirements are so environment- and
business-dependent that for many applications it may be problematic to force people
to make hard decisions in advance for each possible combination of circumstances.
One approach may be to specify goals (making sure that they make sense) and only
then, if they are not met, make the decision what to do with all the information
available.
21

Determining what specific performance requirements is another large topic
that is difficult to formalize. Consider the approach suggested by Peter Sevcik forthat is difficult to formalize. Consider the approach suggested by Peter Sevcik for
finding T, the threshold between satisfied and tolerating users. T is the main
parameter of the Apdex (Application Performance Index) methodology, providing a
single metric of user satisfaction with the performance of enterprise applications.
Peter Sevcik defined ten different methods [Sevcik08].
22

•Default value (the Apdex methodology suggest 4 sec)
•Empirical data•Empirical data
•User behavior model (number of elements viewed / task repetitiveness)
•Outside references
•Observing the user
23

•Controlled performance experiment
•Best time multiple•Best time multiple
•Find frustration threshold F first and calculate T from F (the Apdex methodology
assumes that F = 4T)
•Interview stakeholders
•Mathematical inflection point.
Each method is discussed in details in [Sevcik08].
24

The idea is use of several (say, three) of these methods for the same system.
If all come to approximately the same number, they give us T. While the approachIf all come to approximately the same number, they give us T. While the approach
was developed for production monitoring, there is definitely a strong correlation
between T and the response time goal (having all users satisfied sounds a pretty
good goal), and between F and the response time requirement. So the approach
probably can be used for getting response time requirements with minimal
modifications. While some specific assumptions like four seconds for default or the
F=4T relationship may be up for argument, the approach itself conveys the
important message that there are many ways to determine a specific performance
requirement and it would be better to get it from several sources for validation
purposes. Depending on your system, you can determine which methods from the
above list (or maybe some others) are applicable, calculate the metrics and
determine your requirements.
25

Usability requirements, mainly related to response times, are based on the
basic principles of human-computer interaction. Many researchers agree that usersbasic principles of human-computer interaction. Many researchers agree that users
lose focus if response times are more than eight to 10 seconds and that response
times should be two to five seconds for maximum productivity. These usability
considerations may influence design choices (such as using several Web pages
instead of one). In some cases, usability requirements are linked closely to business
requirements; for example, make sure that your system's response times are not
worse than response times of similar or competitor systems.
26

As long ago as 1968, Robert Miller's paper "Response Time in Man-
Computer Conversational Transactions" described three threshold levels of human
27
Computer Conversational Transactions" described three threshold levels of human
attention [Miller68]. Jakob Nielsen believes that Miller’s guidelines are
fundamental for human-computer interaction, so they are still valid and not likely to
change with whatever technology comes next [Nielsen94]. These three thresholds
are:
• Users view response time as instantaneous (0.1-0.2 second)
• Users feel they are interacting freely with the information (1-5 seconds)
• Users are focused on the dialog (5-10 seconds)

Users view response time as instantaneous (0.1-0.2 second): Users feel that
they directly manipulate objects in the user interface. For example, the time from
28
they directly manipulate objects in the user interface. For example, the time from
the moment the user selects a column in a table until that column highlights or the
time between typing a symbol and its appearance on the screen. Robert Miller
reported that threshold as 0.1 seconds. According to Peter Bickford 0.2 second
forms the mental boundary between events that seem to happen together and those
that appear as echoes of each other [Bickford97].
Although it is a quite important threshold, it is often beyond the reach of
application developers. That kind of interaction is provided by operating system,
browser, or interface libraries, and usually happens on the client side, without
interaction with servers (except for dumb terminals, that is rather an exception for
business systems today).

Users feel they are interacting freely with the information (1-5 seconds):
They notice the delay, but feel the computer is "working" on the command. The
29
They notice the delay, but feel the computer is "working" on the command. The
user's flow of thought stays uninterrupted. Robert Miller reported this threshold as
one-two seconds [Miller68].
Peter Sevcik identified two key factors impacting this threshold [Sevcik03]:
the number of elements viewed and the repetitiveness of the task. The number of
elements viewed is, for example, the number of items, fields, or paragraphs the user
looks at. The amount of time the user is willing to wait appears to be a function of
the perceived complexity of the request. The complexity of the user interface and
the number of elements on the screen both impact thresholds. Back in 1960s
through 1980s the terminal interface was rather simple and a typical task was data
entry, often one element at a time. So most earlier researchers reported that one to
two seconds was the threshold to keep maximal productivity. Modern complex user
interfaces with many elements may have higher response times without adversely
impacting user productivity. Users also interact with applications at a certain pace
depending on how repetitive each task is. Some are highly repetitive; others require
the user to think and make choices before proceeding to the next screen. The more
repetitive the task, the better expected response time.
That is the threshold that gives us response time usability goals for most
user-interactive applications. Response times above this threshold degrade
productivity. Exact numbers depend on many difficult-to-formalize factors, such as
the number and types of elements viewed or repetitiveness of the task, but a goal of
three to five seconds is reasonable for most typical business applications.

There are researches suggests that response time expectations increase with
time. Forrester research [Forrester09] of 2009 suggests 2 second response time, intime. Forrester research [Forrester09] of 2009 suggests 2 second response time, in
2006 similar research suggested 4 seconds (both researches were sponsored by
Akamai, a provider of Web accelerating solutions). While the trend probably exists,
the approach of this research was often questioned: they just ask. It is known that
user perception of time may be misleading. Also, as mentioned earlier, response
time expectations depends on the number of elements viewed, the repetitiveness of
the task, user assumptions of what the system is doing, and UI showing the status.
Stating standard without specification of what page we are talking about may be
overgeneralization.
30

Users are focused on the dialog (5-10 seconds): They keep their attention on
the task. Robert Miller reported that threshold as 10 seconds [Miller68]. Users will
31
the task. Robert Miller reported that threshold as 10 seconds [Miller68]. Users will
probably need to reorient themselves when they return to the task after a delay
above this threshold, so productivity suffers.
Peter Bickford investigated user reactions when, after 27 almost
instantaneous responses, there was a two-minute wait loop for the 28th time for the
same operation. It took only 8.5 seconds for half the subjects to either walk out or
hit the reboot [Bickford97]. Switching to a watch cursor during the wait delayed the
subject's departure for about 20 seconds. An animated watch cursor was good for
more than a minute, and a progress bar kept users waiting until the end. Bickford's
results were widely used for setting response times requirements for Web
applications.
That is the threshold that gives us response time usability requirements for
most user-interactive applications. Response times above this threshold cause users
to lose focus and lead to frustration. Exact numbers vary significantly depending on
the interface used, but it looks like response times should not be more than eight to
10 seconds in most cases. Still, the threshold shouldn’t be applied blindly; in many
cases, significantly higher response times may be acceptable when appropriate user
interface is implemented to alleviate the problem.

The third category, technological requirements, comes from chosen design
and used technology. Some of technological requirements may be known from theand used technology. Some of technological requirements may be known from the
beginning if some design elements are given, but others are derived from business
and usability requirements throughout the design process and depend on the chosen
design.
For example, if we need to call ten Web services sequentially to show the
Web page with a three-second response time, the sum of response times of each
Web service, the time to create the Web page, transfer it through the network and
render it in a browser should be below 3 second. That may be translated into
response time requirements of 200-250 milliseconds for each Web service. The
more we know, the more accurately we can apportion overall response time to Web
services.
Another example of technological requirements is resource consumption
requirements. For example, CPU and memory utilization should be below 70% for
the chosen hardware configuration.
33

Business requirements should be elaborated during design and development,
and merge together with usability and technological requirements into the finaland merge together with usability and technological requirements into the final
performance requirements, which can be verified during testing and monitored in
production. The main reason why we separate these categories is to understand
where the requirement comes from: is it a fundamental business requirement and the
system fails if we miss it or a result of a design decision that may be changed if
necessary.
Performance requirements are important input for Software Performance
Engineering [Smith02]. During design and development the requirements are
further elaborated. For example, the service / stored procedure response time
requirements should be determined by its share in the end-to-end performance
budget. In this way, the worst possible combination of all required services,
middleware and presentation layer overheads will provide the required time. For
example, if there is a Web page with 10 drop-down boxes calling 10 separate
services, the response time objective for each service may be 0.2 seconds to get
three seconds average response time (leaving one second for network, presentation,
and rendering).
34

Requirement Engineering / Architect’s vocabulary is very different from
what is used in performance testing / capacity planning. Performance and scalabilitywhat is used in performance testing / capacity planning. Performance and scalability
are examples of Quality Attributes (QA), part of Nonfunctional Requirements
(NFR).
In addition to specifying requirements in plain text, there are multiple
approaches to formalize documenting of requirements. For example, Quality
Attribute Scenarios by The Carnegie Mellon Software Engineering Institute (SEI) or
Planguage (Planning Language) introduced by Tom Gilb.
35

QA scenario defines source, stimulus, environment, artifact, response, and
response measure [Bass03]. For example, the scenario may be “Users initiate 1,000response measure [Bass03]. For example, the scenario may be “Users initiate 1,000
transactions per minute stochastically under normal operations, and these
transactions are processed with an average latency of two seconds.“ For this
example:
•Source is a collection of users.
•Stimulus is the stochastic initiation of 1,000 transactions per minute.
•Artifact is always the system's services.
•Environment is the system state, normal mode in our example.
•Response is processing the transactions.
•Response measure is the time it takes to process the arriving events (an average
latency of two seconds in our example).
36

Planguage (Planning language) was suggested by Tom Gilb and may work
better for quantifying quality requirements [Simmons01]. Planguage keywordsbetter for quantifying quality requirements [Simmons01]. Planguage keywords
include:
• Tag: a unique identifier
• Gist: a short description
• Stakeholder: a party materially affected by the requirement
• Scale: the scale of measure used to quantify the statement
• Meter: the process or device used to establish location on a Scale
• Must: the minimum level required to avoid failure
• Plan: the level at which good success can be claimed
• Stretch: a stretch goal if everything goes perfectly
• Wish: a desirable level of achievement that may not be attainable through
available means
• Past: an expression of previous results for comparison
• Trend: an historical range or extrapolation of data
• Record: the best-known achievement
It is very interesting that Planguage defines four levels for each requirement:
minimum, plan, stretch, and wish.
37

Another question is how to specify response time requirements or goals.
Individual transaction response times vary, so aggregate values should be used. For
38
Individual transaction response times vary, so aggregate values should be used. For
example, such metrics as average, maximum, different kinds of percentiles, or
median. The problem is that whatever aggregate value you use, you lose some
information.

Percentiles are more typical in SLAs (Service Level Agreements). For
example, "99.5 percent of all transactions should have a response time less than fiveexample, "99.5 percent of all transactions should have a response time less than five
seconds". While that may be sufficient for most systems, it doesn’t answer all
questions. What happens with the remaining 0.5 percent? Do these 0.5 percent of
transactions finish in six to seven seconds or do all of them timeout? You may need
to specify a combination of requirements: for example, 80 percent below four
seconds, 99.5 percent below six seconds, and 99.9 percent below 15 seconds
(especially if we know that the difference in performance is defined by distribution
of underlying data). Other examples may be average four seconds and maximal 12
seconds, or average four seconds and 99 percent below 10 seconds.
39

As Adrian Cockcroft noted about observability [Cockcroft00], in addition to
losing information when you aggregate, there are different viewpoints forlosing information when you aggregate, there are different viewpoints for
performance data that need to be provided for different audiences. You need
different metrics for management, engineering, operations, and quality assurance.
For operations and management percentiles may work best. If you do performance
tuning and want to compare two different runs, average may be a better metric to
see the trend. For design and development you may need to provide more detailed
metrics; for example, if the order processing time depends on the number of items
in the order, it may be separate response time metrics for one to two, three to 10, 10
to 50, and more than 50 items.
Moreover, the tools providing performance information for different
audiences are usually different, present information in a different way, and may
measure different things. For example, load testing tools and active monitoring tools
provide metrics for the used synthetic workload that may differ significantly from
the actual production load. This becomes a real issue if you want to implement
some kind of process, such as Six Sigma, to keep performance under control
throughout the whole system lifecycle.
40

Things get more complicated when there are many different types of
transactions, but a combination of percentile-based performance and availabilitytransactions, but a combination of percentile-based performance and availability
metrics usually works in production for most interactive systems. While more
sophisticated metrics may be necessary for some systems, in most cases they make
the process overcomplicated and results difficult to analyze.
There are efforts to make an objective user satisfaction metric. For example,
Apdex - Application Performance Index [Apdex]. Apdex is a single metric of user
satisfaction with the performance of enterprise applications. The Apdex metric is a
number between 0 and 1, where 0 means that no users were satisfied, and 1 means
all users were satisfied. The approach introduces three groups of users: satisfied,
tolerating, and frustrated. Two major parameters are introduced: threshold response
times between satisfied and tolerating users T, and between tolerating and frustrated
users F [Apdex, Sevcik08]. There probably is a relationship between T and the
response time goal, and between F and the response time requirement. However,
while Apdex is a good metric for management and operations, it may be too high-
level for engineering.
41

Requirements validation is making sure that requirements are valid.
Unfortunately term ‘validation’ is quite often used to mean checking against testUnfortunately term ‘validation’ is quite often used to mean checking against test
results instead of verification.
A good way to validate a requirement is too get it from different
independent sources: if all numbers are about the same, it is a good indication that
the requirement is probably valid. Validation may include, for example, reviews,
modeling, and prototyping. Requirements process is iterative by nature and
requirements may change with time, so it is important to trace requirements back to
their source.
43

Requirements verification is checking if the system performs according to
the requirements. To make meaningful comparison, both the requirements andthe requirements. To make meaningful comparison, both the requirements and
results should use the same aggregates. One consideration here is that load testing
tools and many monitoring tools measure only server and network time. While end
user response times, which business is interested in and usually assumed in
performance requirements, may differ significantly, especially for rich web clients
or thick clients due to client-side processing and browser rendering. Verification
should be done using load testing results as well as during ongoing production
monitoring. Checking production monitoring results against requirements and load
testing results is also a way to validate that load testing was done properly.
44

Requirement verification presents another subtle issue: how to differentiate
performance issues from functional bugs exposed under load. Often, additionalperformance issues from functional bugs exposed under load. Often, additional
investigation is required before you can determine the cause of your observed
results. Small anomalies from expected behavior are often signs of bigger problems,
and you should at least figure out why you get them.
When 99 percent of your response times are three to five seconds (with the
requirement of five seconds) and 1 percent of your response times are five to eight
seconds it usually is not a problem. But it probably should be investigated if this 1
percent fail or have strangely high response times (for example, more than 30 sec)
in an unrestricted, isolated test environment. This is not due to some kind of
artificial requirement, but is an indication of an anomaly in system behavior or test
configuration. This situation often is analyzed from a requirements point of view,
but it shouldn't be, at least until the reasons for that behavior become clear.
45

These two situations look similar, but are completely different in nature:
46
1) The system is missing a requirement, but results are consistent: this is a
business decision, such as a cost vs. response time trade off.
2) Results are not consistent (while requirements can even be met): that may
indicate a problem, but its scale isn't clear until investigated.
Unfortunately, this view is rarely shared by development teams too eager to
finish the project, move it into production, and move on to the next project. Most
developers are not very excited by the prospect of debugging code for small
memory leaks or hunting for a rare error that is difficult to reproduce. So the
development team becomes very creative in finding "explanations". For example,
growing memory and periodic long-running transactions in Java are often explained
as a garbage collection issue. That is false in most cases. Even in the few cases
when it is true, it makes sense to tune garbage collection and prove that the problem
went away.
Another typical situation is getting some transactions failed during
performance testing. It may still satisfy performance requirements, which, for
example, state that 99% of transactions should be below X seconds – and the share
of failed transaction is less than 1 percent. While this requirement definitely makes
sense in production where we may have network and hardware failures, it is not
clear why we get failed transactions during the performance test if it was run in a
controlled environment and no system failures were observed. It may be a bug
exposed under load or a functional problem for some combination of data.

When some transactions fail under load or have very long response times in
the controlled environment and we don't know why, we have one or more problems.the controlled environment and we don't know why, we have one or more problems.
When we have unknown problems, why not trace it down and fix in the controlled
environment? It would be much more difficult in production. What if these few
failed transactions are a view page for your largest customer, and you won't be able
to create any order for this customer until the problem is fixed? In functional
testing, as soon as you find a problem, you usually can figure out how serious it is.
This is not the case for performance testing: usually you have no idea what caused
the observed symptoms and how serious it is, and quite often the original
explanations turn out to be wrong.
Michael Bolton described this situation concisely [Bolton06]:
As Richard Feynman said in his appendix to the Rogers Commission Report
on the Challenger space shuttle accident, when something is not what the design
expected, it's a warning that something is wrong. "The equipment is not operating
as expected, and therefore there is a danger that it can operate with even wider
deviations in this unexpected and not thoroughly understood way. The fact that this
danger did not lead to a catastrophe before is no guarantee that it will not the next
time, unless it is completely understood." When a system is in an unpredicted state,
it's also in an unpredictable state.
47

We need to specify performance requirements at the beginning of any
project for design and development (and, of course, reuse them during performance
48
project for design and development (and, of course, reuse them during performance
testing and production monitoring). While performance requirements are often not
perfect, forcing stakeholders just to think about performance increases the chances
of project success.
What exactly should be specified – goal vs. requirements (or both), average
vs. percentile vs. APDEX, etc. – depends on the system and environment. Whatever
it is, it should be something quantitative and measurable in the end. Making
requirements too complicated may hurt. We need to find meaningful goals /
requirements, not invent something just to satisfy a bureaucratic process.
If we define a performance goal as a point of reference, we can use it
throughout the whole development cycle and testing process and track our progress
from the performance engineering point of view. Tracing this metric in production
will give us valuable feedback that can be used for future system releases.

References
[Apdex] Apdex Web site[Apdex] Apdex Web site
http://www.apdex.org/
[Barber07] Barber, S. Get performance requirements right - think like a user, Compuware white paper, 2007.
http://www.perftestplus.com/resources/requirements_with_compuware.pdf
[Bass03] Bass L., Clements P., Kazman R. Software Architecture in Practice, Addison-Wesley, 2003.
http://etutorials.org/Programming/Software+architecture+in+practice,+second+edition
[Bickford97] Bickford P. Worth the Wait? Human Interface Online, View Source, 10/1997.
http://web.archive.org/web/20040913083444/http://developer.netscape.com/viewsource/bickford_wait.htm
[Bolton06] Bolton M. More Stress, Less Distress, Better Software, November 2006.
http://www.stickyminds.com/sitewide.asp?ObjectId=11536&Function=edetail&ObjectType=ART
[Cockcroft00] Cockcroft, A., Walker B. Capacity Planning for Internet Services. Quick planning techniques for high growth rates. Sun
Microsystems, 2000.
http://java.coe.psu.ac.th/SunDocuments/SunBluePrints/caphi.pdf
[Forrester09] eCommerce Web Site Performance Today. Forrester Consulting on behalf of Akamai Technologies, 2009.
http://www.akamai.com/html/about/press/releases/2009/press_091409.html
[Miler68] Miller, R. B. Response time in user-system conversational transactions, In Proceedings of the AFIPS Fall Joint Computer
Conference, 33, 1968, 267-277.
[Microsoft10] Mailbox Server Processor Capacity Planning. Microsoft, 2010.
http://technet.microsoft.com/en-us/library/ee712771.aspx
[Martin86] Martin, G.L. and Corl, K.G. System response time effects on user productivity, Behavior and Information Technology, 5(1), 1986,
3-13.
[Nielsen94] Nielsen J. Response Times: The Three Important Limits, Excerpt from Chapter 5 of Usability Engineering, 1994.
http://www.useit.com/papers/responsetime.html
[Performance07] Performance Testing Guidance for Web Applications, 2007.
http://perftestingguide.codeplex.com/releases/view/6690
[Podelko07] Podelko A. Multiple Dimensions of Performance Requirements, CMG, 2007.
http://www.alexanderpodelko.com/docs/Performance_Requirements_CMG07.pdf
[Sevcik03] Sevcik, P. How Fast Is Fast Enough, Business Communications Review, March 2003, 8–9.
http://www.bcr.com/architecture/network_forecasts%10sevcik/how_fast_is_fast_enough?_20030315225.htm
[Sevcik08] Sevcik, P. Using Apdex to Manage Performance, CMG, 2008.
http://www.apdex.org/documents/Session318.0Sevcik.pdf
[Simmons01] Simmons E. Quantifying Quality Requirements Using Planguage, Quality Week, 2001.
http://www.clearspecs.com/downloads/ClearSpecs20V01_Quantifying%20Quality%20Requirements.pdf
[Smith02] Smith C., Williams L. “Performance Solutions”, Addison-Wesley, 2002.
[SWEBOK04] Guide to the Software Engineering Body of Knowledge (SWEBOK). IEEE, 2004.
http://www.computer.org/portal/web/swebok
49

Performance Requirements: CMG'11 slides with notes (pdf)

More Related Content

Similar to Performance Requirements: CMG'11 slides with notes (pdf) (20)

More from Alexander Podelko (15)

Recently uploaded (20)

Performance Requirements: CMG'11 slides with notes (pdf)