KP3052 Network

1
KP3052 Network System Performance

• Introduction to Network Performance
• Why predict network performance?

2
Network Performance

• Networks are built to support business
  applications
• word processing
• computer-aided drawing
• electronic mail
• imaging etc

3

• Critical Success Factor - each business
  application must be performed in real time
• Real Time - definition depends on the application
  and can vary from mSecs to hours

4

• What Real Time Responses are Expected to satisfy
  the following Business Applications ?
• Word Processing
• Computer aided drawing
• Electronic mail

5
Real Time Response Expected

• Word Processing
• user loads document, makes changes, writes back
  to file server
• read/write operations should be a few seconds at
  most
• Computer aided drawing
• drawing activities are iterative
• users expect subsecond response time when loading
  a library symbol or deleting a component from a
  drawing

6

• Electronic mail
• may take minutes or hours without users becoming
  upset with network performance

7

• How is User Productivity affected by poor
  Response Time?

8
Real Time Response Expected

• User Productivity
• Declines if applications do not respond quickly
  enough, as User
• frustration sets in
• attitudes deteriorate
• work output slows down
• confidence in the network becomes low

9
Real Time Response Expected

• Network Performance must be sufficient to support
  the users/applications response time
  expectations

10
Common Performance Metrics

• User Response Time
• determined by adding up all the delays that
  application packets endure before a response is
  returned to user
• delays - client side think time,
• network delays in forwarding packets,
  
• server side processing time etc
• Two Definitions

11
Common Performance Metrics

• Network Utilization
• current use of available bandwidth, 10, 20 etc
• ability to provide additional bandwidth on demand
• ability to provide dedicated bandwidth to an
  application

12
Network Performance

• good performance must be available day and night
• response time is provided when needed most
• end of fiscal quarter
• end of month
• 9am - 11am each morning to handle Peak
  Interactive Traffic Load etc

13
Network Performance

• good performance must be maintained when
• additional users (remote or local) added to
  network
• new department moves onto network
• new application is deployed enterprise wide
• additional network printers installed etc

14
Additional Users

• If additional users (remote or local) are to be
  added to network
• need to anticipate performance impact in advance
• if analysis indicates
• extra traffic can be carried -gt proceed
• response time will degrade -gt delay until
  selection and implementation of appropriate
  network upgrades

15
Network Performance

• Ability to predict network performance allows us
  to
• Supply adequate network bandwidth
• Keep user response time low
• Increase user productivity
• Provide for future growth
• Ensure successful deployment of new applications

16
Network Performance

• Validate response time goals of new network
  designs
• Troubleshoot for bottlenecks
• Choose among several competing network
  applications
• Choose the best alternative network topology

17

• What are the Costs of a Poorly Performing Network
  ?

18
Costs of a Poorly Performing Network

• Putting a dollar figure on network down time is
  an exercise in estimation
• A Network is considered to be down when users
  cannot work

19
Costs of a Poorly Performing Network

• If the network is down
• A salesperson writing a quotation for a customer
  risks losing business if the quotation cannot be
  delivered on time
• A point-of-sale clerk who cannot authorize a
  customers credit card will lose a sale and
  possibly a customer
• Electronic mail between a contractor and a
  supplier is delayed and a critical order date is
  missed

20
Costs of a Poorly Performing Network

• For many companies the enterprise network (their
  computer systems and networks) has become Mission
  Critical.
• That means, failure or poor performance means
  that the company is out of business

21
Costs of a Poorly Performing Network A
corporate case study

• A Large Australian company with offices across
  all states and New Zealand
• Has central corporate office and numerous
  branches and manufacturing sites interconnected
  by WAN
• Branch offices
• provide sales and support services to customers
• transmit sales orders to the corporate office
• communicate with manufacturing sites for product
  support

22
(No Transcript)
23
A corporate case study

• Corporate office
• communicates with appropriate manufacturing sites
  to schedule product shipments to customers
• The enterprise network is Mission Critical to
  such a company.
• Performance problems can cost dollars in lost
  sales.
• Now think of some other mission critical systems
  ?

24
Costs of a Poorly Performing Network

• Productivity - Loss Estimate
• Full-time administrative assistant 60 per hr
  includes on-costs
• 10 assistants
• Access to File server is slow in the afternoons
  gt 10 of an assistants time is unproductive
• Estimate
• 48weeks x 5days/week x 8hrs/day 1920 hrs
• 0.10 x 1920 x 60 x 10 115,200 lost annually

25
Costs of a Poorly Performing Network

• Software Company - pending product release
• 40 developers and testers
• 7 development systems acting as repository for
  source code, each equipped with software
  development tools
• Periodical build is done, all source code and
  data files are compiled and linked to produce
  final product

26

• Productivity - Loss Estimate
• Turnaround time to be competitive these builds
  should take less than 12 hrs so testing can
  resume next day
• If builds start to take longer, the next release
  will slip because developers cannot test until
  later next day

27
Costs of a Poorly Performing Network

• Productivity - Loss Estimate
• Assume a new release will generate a 1 million
  in revenue over 6 months
• A slip of 1 month represents a direct loss of
  167,000
• A slip in release date may also mean a competitor
  takes market share
• The loss of productivity due to a late delivery,
  missed deadline, or an inoperative application is
  just as real, though not easily estimated in
  dollar terms

28
Costs of a Poorly Performing Network

• Lost sales opportunities
• Low customer and user satisfaction
• Slipped schedules
• Low user morale
• Loss of market share
• Out of business

29

• Performance Studies

30
Performance Studies

• It is often useful to be able to
• predict the affect on a Systems Performance, of
  a change to the system or change in the workload
• determine the minimum network resources required
  to meet an applications requirements

31

• Ask Questions like -
• How is Response Time (R/T) affected ?
• How is Transaction Throughput affected ?
• Where is the bottleneck ?
• Will one 64kbps link be sufficient ?

32
Performance Studies

• Common Approaches
• 1) Do an After-the-Fact Study
• 2) Make a Simple Projection
• 3) Develop an Analytical Model
• 4) Program and run a Simulation Model

33
1) Do an After-the-Fact Study

• Basically you make the changes then Wait-and-See
  what happens
• This tends to lead to
• unhappy clients, ie average R/T goes from 3secs
  to 30secs
• unwise purchases being made

34
1) Do an After-the-Fact Study

• In effect, it is no option at all
• It is also a way of ensuring that you have a
  short but (in) glorious career

35
2) Make a Simple Projection

• Make a simple projection based on current
  knowledge of the Systems behaviour
• For example we could assume a Straight line
  (y2x1) relationship between R/T (y) and System
  load (x) based on the behaviour of the current
  system up to a transaction load of 500 trans/sec
• Refer Diagram

36
(No Transcript)
37

• However what if there is an Exponential (y2x21)
  relationship ? R/T collapses between 0.8 and 0.9
  system load

38
3) Develop an Analytical Model

• An Analytical Model can be expressed as a Set of
  Equations that can be solved to yield a set of
  estimated results that allow us to predict the
  affect of System Change on
• Response Time
• Transaction Throughput
• Number of communications links required
• etc

39

• For Distributed Computing Systems which involve,
  shared facilities - networks, databases,
  printers, routers - Analytical Models based on
  Queuing Theory provide a reasonable fit to
  reality

40
4) Program and run a Simulation Model

• Simulation language (Simscript) or simulation
  package (OPNET) can model reality in great detail
• Overheads
• Time to build Model
• Time to run simulation
• Time to analyse results

41

• Due to development of (cheap) powerful CPUs
  simulation now often used in training design
• Flight simulators
• Architectural design, walk through the proposed
  building
• Virtual Reality environments

42
Performance Studies - Telephone Industry

• Telephone industry pioneered the study of voice
  traffic patterns in order to design and provision
  a profitable circuit switching network
• Engineers can
• Over Provision design a non-blocking switching
  system that guarantees all users can get a dial
  tone anytime. Cost high, user fees high -gt few
  users, low profit
• Under Provision under design, too many users
  will be denied service and again low profit

43

• Performance studies aim to reduce the risk of
  Over or Under provision, hence resulting in a
  network (design) that will provide services most
  of the time

44
Performance Studies

• Risk of Over Provisioning
• Money spent needlessly on hardware and services
  that remain under utilized
• Risk of Under Provisioning
• Users immediately suffer a fixed period of lost
  productivity while additional time and money is
  spent to upgrade the network

45
(No Transcript)
46
Performance studiesNetwork provisioning

• Best reason for a network performance analysis is
  to locate point where provisioning levels
  maximise profit
• The peak is also the low-risk region

47
Performance Studies

• Lets examine some real net performance
  questions
• How does user Response-Time change?
• Should a server be upgraded or a second server
  installed?
• How many PCs can be supported on the a LAN before
  Response-Time deteriorates?
• ?

48

• Can we add a NEW application to the Enterprise
  Network?
• How do we decide WAN Link Speeds

49
How does User Response-Time Change ?

• Situation Resource Utilization is low and there
  are few active users. Examples
• when only a few digital images accessed per hour
• Users do not have to wait
• when there is little new mail to process
• Email system responds quickly
• when there are only a few small reports printed
  per hour
• No waiting for print jobs

50
How does User Response-Time Change ?

• In this situation Response time
• will remain constant
• is independent of the number of userswe can add
  more users without increase in response time

51
How does User Response-Time Change ?

• Situation Utilization High Many active users
• In this situation Response time
• Increases as new users are added
• will not remain constant
• Long queues and lengthy wait timesusers discover
  another way to measure response timetheir
  position in the print queue!

52
How does User Response-Time Change ?
53
How does User Response-Time Change ?

• Queuing theory and simulation methods can provide
  estimates of how response time behaves versus the
  number of users

54
How does User Response-Time Change ?

• Situation Resource Utilization High and Many
  active users
• Given this situation Response Time increases with
  increases in
• the number of active users
• user workload intensity
• the volume of data moved
• queue depth
• level of network utilisation

55
Should a Server be Upgraded or a Second Server
installed ?

• Response time includes delay introduced by server
• As workload on server increases, its utilisation
  approaches 100 and user response time degrades

56
Should a Server be Upgraded or a Second Server
installed ?

• Should we upgrade server by
• increasing number of disks
• installing faster disks
• replacing LAN adaptor with 32 bit version
• adding more RAM
• upgrading CPU
• Etc
• Where is the bottleneck ?

57
Should a Server be Upgraded or a Second Server
installed ?

• Another option is to purchase a second server and
  redistribute the user workload between them. ie.
  Instead of replicating the server, spread the
  workload based on function required.
• Queuing Theory can be used to answer this
  question.

58
Should a Server be Upgraded or a Second Server
installed ?

• Queuing Theory - it is always better to have one
  fast server than two half speed servers
• When the single fast server is run at
• low utilisation, its response time will be almost
  half of the half-speed servers
• high utilisation, its response time will be about
  equal to a half-speed servers
• What other issue may come into consideration ?

59
How many PCs can be supported on the a LAN before
Response Time deteriorates ?

• Analysis (by Rule-of-Thumb)
• PC generates about 1 utilisation
• Keep LAN utilisation below 35 - response time
  becomes erratic when LAN utilisation exceeds 35
• For users response is more important than
  utilisation
• If we know up to 50 of users are active at any
  time, how many PCs can we attach to the LAN ?

60
How many PCs can be supported on the a LAN before
Response Time deteriorates ?

• A LAN analyser or an SNMP management station can
  be used to get a more accurate measure of the
  traffic generated by PCs
• What Media Access Control technique is being used
  ?

61
How many PCs can be supported on the a LAN before
Response Time deteriorates ?

• Max Utilisation Users Active PCs
  Traffic/PC
• Max Utilisation 40/100
• Users Active 50/100
• Traffic per PC 1/100

62
(No Transcript)
63
Can we add a NEW application to the Enterprise
Network ?

• The ability of network to support additional
  traffic comes under scrutiny when new application
  is to be deployed,
• new application and database servers
• additional user workstations
• LAN and WAN traffic will increase.

64

• Analysis 1
• benchmark new application on isolated LAN and
  measure LAN traffic generated by each transaction
• determine location of application and database
  servers
• determine number/type of transactions at each
  location
• Is Utilisation excessive?

65
Can we add a NEW application to the Enterprise
Network ?

• Analysis 2
• determine network topology
• measure existing work traffic flows
• Input topology and traffic details into a
  simulation tool
• perform simulation to provide response time
  figures for each user site and utilisation
  figures for WANs and LANs
• Result - a reproducible documented analysis that
  answers how the new application can be deployed

66
How do we decide WAN Link Speeds ?

• Available Bandwidth
• varies with offered traffic load
• goes to near-zero at saturation, response time
  becomes more unpredictable at this point, ie the
  variation of response time around its mean
  increases dramatically
• Network designers can provide bandwidth to
  achieve a certain level of response time
• they can fine-tune a routers priority queues to
  favour interactive over bulk traffic
• based on packet size
• sometimes known as traffic shaping

67
How do we decide WAN Link Speeds ?

• Need to consider
• network topology
• location of clients and servers
• work habits of the users
• response time throughput requirements
• transaction traffic loads offered by clients and
  servers

68

• Link speeds can be adjusted to meet the
  requirements using techniques such as
• mean-value analysis
• worst-case analysis
• queuing theory, or simulation

69
End of Topic 1