CS622 Page 1

1
3-location Data Network Design

• 3 locations separated by 200 km among pairs.
• Give the new populations below 296 users, design
  the data network.

2
Voice Traffic vs. Data Traffic
3
Data Traffic Statistics

• 20 of internal email, www, DB traffic occurs in
  the busy hour.
• External email arrives evenly during the day.
• Avg. internal email size 60kB. External email
  size 12kB.
• Each url request generates 6 datagrams to server,
  6 datagrams back to client for setup connections,
  a datagram avg. 128B.
• Its http response is 2kB datagram.

4
Data Base Traffic

• Data distributed in 3 servers, one at each site.
• Each employee makes 100 queries and 5 updates.
• Query
• Query first goes to the local server, then go the
  remote server. Will there need to go to third
  server?
• Query packet avg. 800B, response packet avg.
  3500B.
• Probability of data in a server is 1/3. Evenly
  spread.
• Update
• Update packet avg. 6000B, response packet 500B.

5
Cost of Services and Components

• Cost of PCs, workstations, servers not
  considered.
• Routers can handle 2000 datagrams/sec gtgt the
  traffic
• ? Delay can be neglected.

6
Data Network Design Principle 2.2

• Blocking in not important delay is the issue.
• Highly utilized links are not desirable (large
  delay).
• Design Principle 2.2
• In a voice network, highly utilized links can be
  cost-effective, since they exploit the available
  bandwidth to the fullest extent, and when the
  link is given to a connection it receives a high
  grade of service (circuit switch).
• In a data network, highly utilized links are
  terrible since all call traffic using that link
  suffers inordinate delay.

7
Burstiness

• Burstiness peak rate/avg rate
• Two solutions to simultaneous arrive of data
  calls
• Coordination- e.g., token ring. Allow one to be
  servered. But is not good for WAN.Propagation
  delay (p.d.) for 1000 mile ring1000/1860005msTr
  ansmission delay for 1000 bit packet at
  16Mbps1000/16000000 ltlt p.d.
• Queueing or use store and forward (packet
  switching idea).

8
Common Data Rate
The service time for a packet of n bits on a link
of speed S bps is n/S
9
Token Ring vs. Packet Switching
Propagation delay for 1000 mile
ring1000/186,0005.376 ms Transmission delay for
1000 bit packet at 16Mbps1000/16,000,000

0.0625ms For WAN,
token ring protocol is not suitable. A packet
switching network where each link segement
operates independently is a more efficient
design. Packet switching network can be modeled
as a set of queues.
10
M/M/1 Queue

• Link can be modeled as a M/M/1 queue.

11
M/M/1 Average Waiting Time
12
Total Delay (50ms service time)
13
Initial Data Network Design
14
Cost of Initial Design

• Transit router amortized cost 37000.03111/mon
  th
• 64kbps (or D64) internode link 700/month
• 64kbps internet link 1400/month.

15
Traffic in Busy Hour

• 200.2 traffic in busy hour.

16
Design Principles 2.3 2.4

• 2.3 Seek to make a network where all the links
  have a 50 utilization
• 2.4 Seek to make a network where all the links
  have about 50 utilization and as few links as
  possible are underutilized.
• Example
• Question1 How we calculate the delay?
• Question2 For high speed link, can we have high
  utilization?

17
Apply M/M/1 Formula

• Assume 1000 bytes packet (8000 bits).
• Case1 T1 link1.536Mbps, 50 utilization
• Case2 OC-3 link135Mbps, 80 utilization
• Which one has lower delay?

18
Apply M/M/1 Formula

• Case1 T1 link1,536,000 bps, r0.5
  (50utilization)
• 1/m service timepacketsize/transmission
  speed8000/1536000
• T(1/m)/(1-r)(1/(1-r))(1/m)(1/(1-0.5))8000/1.5
  36M10.4 ms.
• Case2 OC-3 link135Mbps, r0.8 (80 utilization)
• T(1/(1-0.8))(8000/135M)5(8000/135M)2.96ms
• We may be willing tolerate a higher utilization
  on these links. But their delay is quite
  unstable.

19
Calculating Internal Email Traffic

• Internal Email relate to the populations of
  source and destination sites. The ration of
  populations among Anagon, Bregen, and Charmes(1,
  4/3, ¾)
• Let x be the volume of internal email from Analog
  to itself.
• Then the traffic from Anagon to Bregen is 4/3 x.
• The traffic from Anagon to Charmes is ¾ x.
• busy hour internal email 100.2600008296/3600(
  s)78933 bps.
• Counting all directional internal email traffic
  9.507x78933bps ? x8303bps

20
Tabular Represenation of Internal Email Traffic
21
External Email

• In the initial design, each site has its own
  Internet connection. Therefore the external
  emails does not go through inter-site internal
  network.
• Internet links are expensive first targets to
  removethen external emails could go over
  inter-site network.
• With 4000 emails/day, 12000 B/emaileach user
  gets 4000120008/(36008hr296)45.045bpssends
  same 45.045bps external emails.
• Multiply the population in each site we get the
  following external traffic table.

22
Tabular Represenation of External Email Traffic

• 45.045964324.32 bps

23
Busy Hour WWW Traffic

• Outbound small requests traffic
  40fetch/day0.26req/fetch128B/req8b/B/(3600s)
  13.653bps
• Inbound big www document and response
  traffic400.2(6x1282000)8/(3600)49.209bps
• For Anagon, Inbound WWW traffic
  13.653bps961310.72bpsoutbound WWW
  traffic49.209bps964724.05bps

24
Busy Hour WWW Traffic
25
DB Query Flow

• Assume query can be answered by a single remote
  server.

26
Busy Hour DB Traffic

• DB Query Traffic
• 1/3 queries to each remote server500.28008(1
  /3)/36005.930 bps
• Their requests come back500.235008(1/3)/3600
  25.926bps
• DB Update Traffic
• 1/3 updates to each remote server50.260008(1
  /3)/36004.444 bps
• 1/3 updates responses back from each remote
  server50.25008(1/3)/36000.370 bps
• DB Traffic From Anagon to Bregen
• Consider just DB Query(text) 965.93012825.926
  3887.8
• Consider all DB traffic The update traffic
  should not be ignored 96(5.9304.444)128(25.92
  60.370)4357.568

27
DB Traffic Table
28
Busy Hour Traffic (64kbps links)
29
Busy Hour Traffic
30
Drop Algorithm for Network Design

• Drop the lightest utilized component in the
  network.
• Calculate the new routes for all traffic that use
  the dropped component.
• But do we really have control over the routing in
  the network?
• We will examine 3 types of routing
• SNA (IBM System Network Architecture) tight
  control
• OPSF (Open Shortest Path First) some
  control
• RIP (Routing Information Protocol)
  no control

31
Routing in SNA

• On IBM SNA (System Network Architecture),
  designer has up to 16 routes that can be
  specified between a pair of nodes. The paths are
  directional. The return path of a route can go
  through different links.
• Advantage flexible, a lot of control
• Disadvantage adding a node is not automatic,
  required offline programs to generate the paths.

32
OPSF Routing

• Assign each link a length (or weight) in each
  direction.
• Routes are calculated using shortest path
  algorithm.
• Traffic are directed to the next link along the
  shortest path.
• Two routes between a pair of nodes. (compared to
  max. of 16 for SNA)
• Weight can be measured as delay on the
  directional link.
• Link weights can be broadcast periodically and
  routing table recalculated.

33
Routing Information Protocol

• Use hop count instead of accumulated link weight
  for compute the route.
• Does not consider the bandwidth of each link.
• For 1000-byte packet,
• a two hop path with T1 link has(10008b/1.535Mbps
  )210.42ms.
• A single hop path with 9.6kbps link
  has10008b/9600bps833ms.

34
Assumptions for Drop Algorithm

• Assume we can use shortest path routing within
  BMI corp. domain.
• All three inter-site links have a length of 10.
• The distance to all external domains is the same
  through all three gateways.
• Try to reduce cost by removing links and see if
  remaining network remain feasible.

35
Drop Algorithm

• Initially, mark all links as being deletable.
• Find the most expensive deletable link. If there
  is a tie, take the link with the lowest
  utilization. We call this the candidate link for
  deletion.
• If such link exists, delete the link and see if
  the remaining network is feasible (can carry the
  traffic).
• If it is feasible, go back to step 2.
• If not feasible, mark the link not being
  deletable and loop back to step 2.
• If such link does not exist, terminate.

36
Modified Drop Algorithm Code
Consider increase other links capacity
37
Apply Drop Algorithm on Initial Design

• Round 1.
• Step2. Among 3 external links, choose Charmes to
  gateC.
• Step3. Redirect traffic to Gateway A (with less
  traffic)by reducing the length btw Anagon and
  Charmes to 9.
• GateC?Charmes traffic (WWWExternal Email) go
  over GateA?Anagon?Charmes.
• Charmes?GateC traffic go over Charmes?Anagon?Gate
  A
• The new traffic flow is shown next page.

38
Traffic Flow After Removing Link to GateC
All link utilizations lt 0.5 cost saving1400
39
Apply Drop Algorithm on Initial Design

• Round 2.
• Step2. Among 2 external links, choose Bregen to
  gateB since it has less traffic now.
• Step3. Redirect traffic to Gateway A (with less
  traffic)
• GateB?Bregen traffic (WWWExternal Email) go over
  GateA?Anagon?Bregen.
• Bregen?GateB traffic go over Bregen?Anagon?GateA
• The new traffic flow is shown next page.

40
Traffic Flow After Removing Link To GateB

• All link utilizations lt 0.5 cost saving another
  1400

41
Round 3 Round 4

• Round 3 Try to delete link to gateA and find it
  undeletable.
• Round 4 Among the remaining 3 inter-site links,
  Bregen??Charmes has less utilization (add both
  directional traffic).
• Redirect traffic around Anagon.

42
Traffic Flow After Removing Link btw Bregen and
Charmes

• Utilization between Anagon and Bregen high, need
  add link?

43
Rounds 4, 5, 6

• After removing link btw Charmes and Bregen, we
  need to add capacity to Anagon and Bregen ? no
  cost saving.
• We also lose alternative route (less
  reliability).
• Decide not to remove.
• Same results for link btw Anagon and Charmes, and
  link between Anagon to Bregen.
• Algorithm terminates.

44
Drop Algorithm Result

• 2 internet links removed? cost saving 2800/month

45
Where the Drop Algorithm Went Wrong?

• It chooses Anagon instead of Bregen, which has
  most traffic and largest population.
• This force more traffic onto longer paths.
• Lesson Heuristic algorithms often make mistakes.
• If we choose to locate gateway at Bregen, we
  could remove link btw Anagon and Charmes
• Save 700/month
• Save 102/month by placing terminal routers at
  Anagon and Charmes.
• Final cost 3833-700-1023031/month.

46
Final Design

• Definition 2.3 A benign algorithm is one that
  does no damage to a design. It only improve it or
  leave it alone.
• The drop algorithm is not optimal but is it
  benign?

47
Homework 5
A B C D

• Exercise 2.6. Assume the traffic in Table 2.17
  increases 5 times. Design a low-cost solution for
  the BMI Corporation.
• Exercise 2.8. In figure 2.17, we have a 4-site
  network.Given the traffic matrix, use the drop
  algorithm to redesign the network. Each link is
  limited to 28,000bps in each direction. Further
  assume that each link costs 100 times the length
  of the link. A?D link is 100100.

A B C D
48
Solution to Hw5

• Exercise 2.6. 5 times the original Busy Hour
  Traffic

49
Exercise 2.6

• Traffic between sites gt 32kbps ? can not use
  single 64kbps lines at a 50 utilization (design
  goal/principle).
• According to the pricing info. It does not pay
  to have two 64kbps lines. We should go for the
  256kbps line (same price!)

50
Exercise 2.6

• For Internet links, we need to go for 256kbps
  too.
• Actually the total incoming Internet traffic is
  139375bps and requires two 256kbps lines to be
  lower than the 50 utilization.
• We locate the two Internet links at Anagon and
  Bregen since Charmes traffic is the smallest.
• Charmes Internet traffic needs to be rerouted.
  Since Anagon has lower utilization, it was routed
  to Anagon.
• Anagon?Charmes traffic increases by 33930bps.
• Charmes?Anagon traffic increases by 21130bps.

51
Traffic after Eliminating Charmes Internet Link
52
Try to Eliminate InterSite Links

• The first candidate is the BregenCharmes link.
• However, it will increase the utilization of the
  other two links to more than 50. Actually just
  slightly over.
• If the rule is very strict on not to exceed 50,
  then none of the intersite link can be removed.
• The last table indicates the final low-cost
  solution.

53
Final Low-Cost Solution
Gate B
54
Exercise 2.8

• There are two link costs in this network, 10,000
  and 5,800.
• Let us try to remove the expensive 10,000 links
  first.
• The three expensive links all have the same
  10kbps.
• Let us remove A-B link first and reroute traffic
  via C.

55
Exercise 2.8

• Remove A-D link and reroute the traffic via C. We
  get
• Remove B-D link and reroute the traffic via C.

No further link can be reduced without the
remaining links exceeding 28kbps speed.
56
Final Solution

• The above solution is based on aggregated traffic
  of each link not exceed 28kbp, and not that
  prohibits over 50 utilization (14kbps).
• If we use the latter criteria, than you will
  only able to remove A-C link.