1. 6/17/ 2002EMIS 8392 Maya Petkova1 Chapter 4 Traffic and Cost Generators

2. 6/17/ 2002EMIS 8392 Maya Petkova2 Overview Overview Interest to people doing practical network design.Interest to people doing practical network design. Can not get the actual tariffs and traffic.Can not get the actual tariffs and traffic. “ Practice, practice ! ”“ Practice, practice ! ” Sample scenarios.Sample scenarios.

3. 6/17/ 2002EMIS 8392 Maya Petkova3 War Stories War Stories War story 4.1 – a company has to make a decision about converting to routers at the next board meeting. War story 4.1 – a company has to make a decision about converting to routers at the next board meeting. - meeting is in 3 months - meeting is in 3 months - need to have a capital budget and network - need to have a capital budget and network operating cost operating cost - no actual traffic has been measured - no actual traffic has been measured - there are only user populations by job category - there are only user populations by job category PROBLEM: How to take actual measurements and PROBLEM: How to take actual measurements and meet the schedule. meet the schedule. SOLUTION: Network generator. SOLUTION: Network generator.

4. 6/17/ 2002EMIS 8392 Maya Petkova4 War Stories – Cont’d War Stories – Cont’d War story 4.2 - a large company runs a network composed of hundreds of subnetworks with a major restructuring to be done. War story 4.2 - a large company runs a network composed of hundreds of subnetworks with a major restructuring to be done. - the amount of traffic entering and exiting the - the amount of traffic entering and exiting the network at any connection is known network at any connection is known - no traffic matrix available - no traffic matrix available PROBLEM: Build a traffic matrix that matches the given totals for each site. given totals for each site. SOLUTION: Traffic generator.

5. 6/17/ 2002EMIS 8392 Maya Petkova5 More on the Advantages of Synthesizing Data Real network design is 90% file preparation, traffic and cost generators can reduce that proportion to 50% or even 25%.Real network design is 90% file preparation, traffic and cost generators can reduce that proportion to 50% or even 25%. Creation of unlimited number of interesting design problems.Creation of unlimited number of interesting design problems. Reduces the mass of these problems from MB to KB – big structures created parametrically.Reduces the mass of these problems from MB to KB – big structures created parametrically. No worries about the security and privacy issues associated with real data.No worries about the security and privacy issues associated with real data. Can study a broader range of problems than those that can be gathered off of real systems.Can study a broader range of problems than those that can be gathered off of real systems.

6. 6/17/ 2002EMIS 8392 Maya Petkova6 The Structure of a Network Design Problem Need to place switches or multiplexers at the nodes.Need to place switches or multiplexers at the nodes. Interconnect the above equipment with links.Interconnect the above equipment with links. Traffic needs to be routed.Traffic needs to be routed. Performance needs to be evaluated.Performance needs to be evaluated. Above requires data structures and file formats. Almost all tools have proprietary standards.

7. 6/17/ 2002EMIS 8392 Maya Petkova7 A Small Complete Design Problem A Small Complete Design Problem A simple extensible formatA simple extensible format 5 nodes to be interconnected.5 nodes to be interconnected. All the necessary information to build a network model is provided.All the necessary information to build a network model is provided. Organized as a set of tables.Organized as a set of tables.

8. 6/17/ 2002EMIS 8392 Maya Petkova8 The SITES Table The SITES Table Contains basic information about each site.Contains basic information about each site. Note: importance of the PARENT column.Note: importance of the PARENT column.

9. 6/17/ 2002EMIS 8392 Maya Petkova9 The LINETYPES Table The LINETYPES Table Contains the list of the links to be used in the design.

10. 6/17/ 2002EMIS 8392 Maya Petkova10 The TRAFFIC Table The TRAFFIC Table Contains the flows to be carried by the network.Contains the flows to be carried by the network. Flow units:Flow units: - voice traffic – Erlangs - voice traffic – Erlangs - data traffic - bits/second - data traffic - bits/second - multiplexer design – virtual - multiplexer design – virtual circuit capacity circuit capacity

11. 6/17/ 2002EMIS 8392 Maya Petkova11 The TARIFF Table The TARIFF Table Contains the costs of all the links available for the design.Contains the costs of all the links available for the design. By convention missing costs are set to a very large constant LINFINITYBy convention missing costs are set to a very large constant LINFINITY

12. 6/17/ 2002EMIS 8392 Maya Petkova12 The EQUIPMENT Table The EQUIPMENT Table Specifies the cost and capacity of the equipment to be placed at each site.Specifies the cost and capacity of the equipment to be placed at each site. Specifies the throughput of the equipment and the number of lines that can be terminated.Specifies the throughput of the equipment and the number of lines that can be terminated.

13. 6/17/ 2002EMIS 8392 Maya Petkova13 The EQUIPMENT Table – Cont’d The EQUIPMENT Table – Cont’d Throughput measurementThroughput measurement - for data switches - packets/second - for data switches - packets/second - for voice switches – call setups/second - for voice switches – call setups/second - multiplexers usually set up static circuits and - multiplexers usually set up static circuits and have only a degree constraint have only a degree constraint Most of the cost of the network 90% - 95% is in the links rather then in the nodes. Link errors could be very expensive.Most of the cost of the network 90% - 95% is in the links rather then in the nodes. Link errors could be very expensive. Detailed description in Appendix D.Detailed description in Appendix D.

14. 6/17/ 2002EMIS 8392 Maya Petkova14 The PARMS Table The PARMS Table Contains information to guide the design process

15. 6/17/ 2002EMIS 8392 Maya Petkova15 Changing the Problem Size Suppose instead of 5 we have 100 nodes.Suppose instead of 5 we have 100 nodes. What is the effect on the size of the tables?What is the effect on the size of the tables? - LINETYPES, EQUIPMENT, PARMS tables do not change - LINETYPES, EQUIPMENT, PARMS tables do not change - SITES table grows linearly - SITES table grows linearly - TRAFFIC table may contain entries - TRAFFIC table may contain entries - TARIFF table – number of link costs is ( ) - TARIFF table – number of link costs is ( ) May have 10,000 traffic records, 4950 link costs.May have 10,000 traffic records, 4950 link costs. If created by hand will contain thousands and thousands of errors.If created by hand will contain thousands and thousands of errors.

16. 6/17/ 2002EMIS 8392 Maya Petkova16 The Most Important Lesson from Ch.4 - According to the Author All the tables with over a few hundred entries must be produced automatically.All the tables with over a few hundred entries must be produced automatically. Traffic and cost generators will facilitate the process of producing complete design problems provided that tables SITES, LINETYPES, EQUIPMENT and PARMS are put together.Traffic and cost generators will facilitate the process of producing complete design problems provided that tables SITES, LINETYPES, EQUIPMENT and PARMS are put together.

17. 6/17/ 2002EMIS 8392 Maya Petkova17 The SITES Table for Network Generation The SITES Table for Network Generation Squareworld – sites chosen at random.Squareworld – sites chosen at random. Real life – assembled by interviews, reviews of personnel records, real estate records, examining existing networks, management reports.Real life – assembled by interviews, reviews of personnel records, real estate records, examining existing networks, management reports. Design Principle 4.1.Design Principle 4.1. The first thing a network designer needs to know is the location of the sites to be connected, just as builder needs a survey of a building site. The first thing a network designer needs to know is the location of the sites to be connected, just as builder needs a survey of a building site.

18. 6/17/ 2002EMIS 8392 Maya Petkova18 The SITES Table for Network Generation – Cont’d The SITES Table for Network Generation – Cont’d Four additional columns to the previously discussed SITES table. Four additional columns to the previously discussed SITES table. 1.POPULATION – The number of users at a site or the census population. 2.TRAFOUT – The traffic leaving the site for other sites expressed in the appropriate units. 3.TRAFIN – The traffic entering the site from other sites. 4.LEVEL – The level of the site in the hierarchy.

19. 6/17/ 2002EMIS 8392 Maya Petkova19 Sample SITES Table Sample SITES Table

20. 6/17/ 2002EMIS 8392 Maya Petkova20 Traffic Generators Traffic Generators The job of the traffic generator is to spread the traffic around the network.The job of the traffic generator is to spread the traffic around the network. Various types of network have different units of traffic.Various types of network have different units of traffic. Being able to fill in the traffic matrix for data networks is more important than for voice and multiplexer networks since the quantum of traffic is much smaller and the accounting is sketchier.Being able to fill in the traffic matrix for data networks is more important than for voice and multiplexer networks since the quantum of traffic is much smaller and the accounting is sketchier.

21. 6/17/ 2002EMIS 8392 Maya Petkova21 Traffic Generators - Uniform Traffic Traffic Generators - Uniform Traffic Simplest traffic – for all and we defineSimplest traffic – for all and we define Can give rise to interesting problems - CST (communication spanning tree) problem – find the lowest cost tree to connect n sites, where the cost of each link is proportional to the flow across it. In 1994 C.Palmer and A. Kershenbaum produced a creditable algorithm for solving the CST problem.Can give rise to interesting problems - CST (communication spanning tree) problem – find the lowest cost tree to connect n sites, where the cost of each link is proportional to the flow across it. In 1994 C.Palmer and A. Kershenbaum produced a creditable algorithm for solving the CST problem. Not realistic for many networks.Not realistic for many networks. Better fits the traffic inside the switches.Better fits the traffic inside the switches. Used to test the behavior of algorithms.Used to test the behavior of algorithms.

22. 6/17/ 2002EMIS 8392 Maya Petkova22 Traffic Generators – Random Traffic (Voice&Data) Traffic Generators – Random Traffic (Voice&Data) Suppose we have n sites.Suppose we have n sites. Suppose we specify the max and the min for the traffic in either Erlangs or bits/second.Suppose we specify the max and the min for the traffic in either Erlangs or bits/second. Randreq.c generates random traffic in the (min, max) range between each pair of sites.Randreq.c generates random traffic in the (min, max) range between each pair of sites. for (i=0; i<n; ++i) { for (i=0; i<n; ++i) { for (j=0; j<n; ++j) { for (j=0; j<n; ++j) { traffic[i][j]=min_req+((double)rand())*(max_req- min_req)/RAND_MAX; traffic[i][j]=min_req+((double)rand())*(max_req- min_req)/RAND_MAX; } /* endfor */ } /* endfor */

23. 6/17/ 2002EMIS 8392 Maya Petkova23 Traffic Generators - Random Traffic (MUX traffic) Traffic Generators - Random Traffic (MUX traffic) Randreq2.c generates random set of circuit requirements for the multiplexer design problem.Randreq2.c generates random set of circuit requirements for the multiplexer design problem. Suppose we have n nodes and nreq=m requirements for circuits of type D56 – 56,000 bps links.Suppose we have n nodes and nreq=m requirements for circuits of type D56 – 56,000 bps links. for (i=0; i < nreq; ++i) { for (i=0; i < nreq; ++i) { end1 = rand()%n+1; end1 = rand()%n+1; end2 = rand()%n+1; end2 = rand()%n+1; if (end1==end2) { if (end1==end2) { --i; --i; continue; continue; } /* end if */ } /* end if */ fprintf(fp,”N%d N%d D56\n”, end1, end2); fprintf(fp,”N%d N%d D56\n”, end1, end2); }/* end for */ }/* end for */

24. 6/17/ 2002EMIS 8392 Maya Petkova24 More Realistic Traffic: A Simple Generator More Realistic Traffic: A Simple Generator Uniform and random traffic generators produce nonrealistic traffic (does not match what is actually observed by traffic surveys).Uniform and random traffic generators produce nonrealistic traffic (does not match what is actually observed by traffic surveys). How to produce realistic traffic?How to produce realistic traffic? - determine the factors that could influence - determine the factors that could influence the total traffic between 2 sites. the total traffic between 2 sites. - assume we have 2 nodes and that - assume we have 2 nodes and that have populations and and are apart. have populations and and are apart.

25. 6/17/ 2002EMIS 8392 Maya Petkova25 More Realistic Traffic: A Simple Generator – Cont’d More Realistic Traffic: A Simple Generator – Cont’d Simplest notion of population is the human population. It could be the portion that uses telecommunications (different for voice&data).Simplest notion of population is the human population. It could be the portion that uses telecommunications (different for voice&data). Network designer chooses how to model the site population.Network designer chooses how to model the site population. If population is homogeneous then we can use a model based on physics.If population is homogeneous then we can use a model based on physics.

26. 6/17/ 2002EMIS 8392 Maya Petkova26 More Realistic Traffic: A Simple Generator – Cont’d More Realistic Traffic: A Simple Generator – Cont’d If we chose model based on gravitation attraction: Pop_Power =1, Dist_Power = 2.If we chose model based on gravitation attraction: Pop_Power =1, Dist_Power = 2. If we chose model based on magnetism:If we chose model based on magnetism: Pop_Power =1, Dist_Power = 3. Pop_Power =1, Dist_Power = 3. If site traffic is independent of population: Pop_Power = 0.If site traffic is independent of population: Pop_Power = 0. In all cases is used as a scaling factor to put the traffic into the range we wish.In all cases is used as a scaling factor to put the traffic into the range we wish.

27. 6/17/ 2002EMIS 8392 Maya Petkova27 Even More Realistic Traffic Models Q. What happens if the 2 nodes in the previous models are in the same location? A.Division by 0. Q.What happens if the 2 nodes have populations of 1,000,000 each and the distance between them is 1000 miles? A. The effect of distance is lost. Goal: Make calculation more unit independent.

28. 6/17/ 2002EMIS 8392 Maya Petkova28 Even More Realistic Traffic Models – Cont’d To normalize the calculation we define:To normalize the calculation we define: Define to avoid division by 0.Define to avoid division by 0. Define to avoid all the traffic to and from small nodes to be 0.Define to avoid all the traffic to and from small nodes to be 0. Define traffic byDefine traffic by

29. 6/17/ 2002EMIS 8392 Maya Petkova29 Even More Realistic Traffic Models – Cont’d Setting the scale factorSetting the scale factor - Set Pop_max =1000 - Set Pop_max =1000 - Set Pop_Power = 1 - Set Pop_Power = 1 - Set dist_max = 500 - Set dist_max = 500 - Set Dist_Power = 1 - Set Dist_Power = 1 We want to get 500 bps of traffic between 2 sites of population 500 with distance of 250.We want to get 500 bps of traffic between 2 sites of population 500 with distance of 250. Select = 1000.Select = 1000.

30. 6/17/ 2002EMIS 8392 Maya Petkova30 Traffic Normalization: Uniform Normalization Traffic Normalization: Uniform Normalization Example 4.1. A company has 50 sites linked by 85 E1 lines. The average number of hops in a route is 2.75. Links have an average utilization of 55%. What value of should be chosen to generate the traffic? Example 4.1. A company has 50 sites linked by 85 E1 lines. The average number of hops in a route is 2.75. Links have an average utilization of 55%. What value of should be chosen to generate the traffic? Total traffic is Total traffic is (85x2x2,048,000x0.55)/2.75 = 69.632 Mbps (85x2x2,048,000x0.55)/2.75 = 69.632 Mbps Let where is the traffic to Let where is the traffic to be carried from node to node be carried from node to node Choose = 69,632,000/T Choose = 69,632,000/T Works for all values of Pop_Power and Dist_Power. Works for all values of Pop_Power and Dist_Power.

31. 6/17/ 2002EMIS 8392 Maya Petkova31 Traffic Normalization: Row Normalization Traffic Normalization: Row Normalization Example 4.2. A company has 50 sites linked by a network. At each site we observe the total flow of traffic out to the network,. We wish to normalize the requirements so that the traffic out to the network for each site matches the observations. Example 4.2. A company has 50 sites linked by a network. At each site we observe the total flow of traffic out to the network,. We wish to normalize the requirements so that the traffic out to the network for each site matches the observations. The traffic from node to all other nodes is The traffic from node to all other nodes is We can define We can define

32. 6/17/ 2002EMIS 8392 Maya Petkova32 Traffic Normalization: Row Normalization – Cont’d Traffic Normalization: Row Normalization – Cont’d Normalization will be different at each node.Normalization will be different at each node. Row normalization allows us to match the row sums to the observed flows.Row normalization allows us to match the row sums to the observed flows. Row normalization is as the matrix multiplication of the diagonal matrix with the matrixRow normalization is as the matrix multiplication of the diagonal matrix with the matrix If measurements available only at some of the sites it is reasonable to use the average of these values for the rest of the nodes in the network.If measurements available only at some of the sites it is reasonable to use the average of these values for the rest of the nodes in the network.

33. 6/17/ 2002EMIS 8392 Maya Petkova33 Traffic Normalization: Row and Column Normalization Traffic Normalization: Row and Column Normalization Example 4.3. A company has 50 sites linked by a network. At each site we observe the total flow of traffic out to the network,.We also observe the traffic flow in from the network We wish to normalize the requirements so that the traffic out to the network and in from network at each node matches the observations. Example 4.3. A company has 50 sites linked by a network. At each site we observe the total flow of traffic out to the network,.We also observe the traffic flow in from the network We wish to normalize the requirements so that the traffic out to the network and in from network at each node matches the observations. Theorem 4.1. If, then the normalization problem has no solution. Theorem 4.1. If, then the normalization problem has no solution. Proof: Obvious, argue using sources and sinks. Proof: Obvious, argue using sources and sinks.

34. 6/17/ 2002EMIS 8392 Maya Petkova34 Traffic Normalization: Row and Column Normalization - Cont’d Traffic Normalization: Row and Column Normalization - Cont’d Rowcol.c normalizes the requirements matrix to match row and column sums.Rowcol.c normalizes the requirements matrix to match row and column sums. No guaranteed success, especially when there are too many zeros in the traffic matrixNo guaranteed success, especially when there are too many zeros in the traffic matrix Works well with the traffic generated by our basic population and distance nodes.Works well with the traffic generated by our basic population and distance nodes. Uses row_scale and col_scale to measure how far each row and column are from the desired totals.Uses row_scale and col_scale to measure how far each row and column are from the desired totals. Variable max_scale gives the max value of any of the scaling factorsVariable max_scale gives the max value of any of the scaling factors Iterates until have converged by having max_scale between 0.999 and 1.001Iterates until have converged by having max_scale between 0.999 and 1.001

35. 6/17/ 2002EMIS 8392 Maya Petkova35 The Final Traffic Generator: The Level Matrix Introduced so far traffic generators produced symmetric traffic –Introduced so far traffic generators produced symmetric traffic – Often traffic is asymmetric – terminal traffic, Web browsing, database query.Often traffic is asymmetric – terminal traffic, Web browsing, database query. To allow for asymmetric traffic we introduce the notion of levels and level matrix.To allow for asymmetric traffic we introduce the notion of levels and level matrix. Definition 4.1. A level matrix will be an matrix with entries in. Definition 4.1. A level matrix will be an matrix with entries in.

36. 6/17/ 2002EMIS 8392 Maya Petkova36 The Final Traffic Generator: The Level Matrix – Cont’d Example: Example: - suppose type 1 nodes are requesting Web pages - suppose type 1 nodes are requesting Web pages - suppose type 2 nodes are supplying Web pages - suppose type 2 nodes are supplying Web pages - suppose each incoming page is 3 times larger than the request - suppose each incoming page is 3 times larger than the request - the level matrix is - the level matrix is

37. 6/17/ 2002EMIS 8392 Maya Petkova37 The Final Traffic Generator: The Level Matrix – Cont’d Further modification of the basic network generation formulaFurther modification of the basic network generation formula If is of level 1 and is of level 2If is of level 1 and is of level 2 If is at level 1, thenIf is at level 1, then

38. 6/17/ 2002EMIS 8392 Maya Petkova38 The Final Traffic Generator: The Level Matrix – Cont’d Realistic network traffic is event driven.Realistic network traffic is event driven. There is certain degree of randomness (unexpected forces).There is certain degree of randomness (unexpected forces). Let rand() is a pseudo-random generator that produces strings of numbers in the range [0, RAND_MAX].Let rand() is a pseudo-random generator that produces strings of numbers in the range [0, RAND_MAX]. We model randomness by adding random factorWe model randomness by adding random factor Absolutely, positively the final traffic generator:Absolutely, positively the final traffic generator:

39. 6/17/ 2002EMIS 8392 Maya Petkova39 The Final Traffic Generator: The Level Matrix – Cont’d The final traffic generator can be combined with any of the 3 normalizations.The final traffic generator can be combined with any of the 3 normalizations. ROW/COL normalization may have some problems with convergence.ROW/COL normalization may have some problems with convergence. Example – use the level matrix from slide 36 for the following data:Example – use the level matrix from slide 36 for the following data:

40. 6/17/ 2002EMIS 8392 Maya Petkova40 Traffic Generators and Sensitivity Analysis Traffic Generators and Sensitivity Analysis Useful to create traffic suites rather than just a single set of traffic.Useful to create traffic suites rather than just a single set of traffic. Allow us to assess the capability of the network to respond to a change.Allow us to assess the capability of the network to respond to a change. With network generator easy – add additional load and see if the network still performs correctlyWith network generator easy – add additional load and see if the network still performs correctly Sensing the capacity is important if there is a requirement not to redesign for a period of timeSensing the capacity is important if there is a requirement not to redesign for a period of time If traffic grows 20% per year and if the network begins to falter at 35% - not a good candidate network.If traffic grows 20% per year and if the network begins to falter at 35% - not a good candidate network.