Network analysis and dimensioning course project Alexander Pyattaev March 10, 2014 Email: alexander.pyattaev@tut.fi 1 Project description Your task in this project is to be a network architect. You will get to plan, evaluate, and estimate costs of running a city-scale network. Playing a major operator, you will provide your customers with triple-play services (audio, video broadcasting and data over the same connection), high-speed data center access, as well as cellular coverage. Unlike most operators, however, you will do all of that right, that is using the latest and best of the available analysis and simulation tools to make sure that your customers get the best quality of service at all times. 1.1 Project outcomes Within this project, you will have the opportunity to get familiar with the usage of advanced networking technologies, including MPLS, ATM, SDH, VLAN models, InfiniBand, 3GPP LTE, SIP and many others. You will also get the knowledge needed to put them to good use in a real city-wide network. In addition to simply having fun with new technologies, you will also understand the tools used for planning of large networks, as well as their subsystems, such as data centers etc. We will cover concepts of redundancy, availability and guaranteed service. 1.2 Project structure The project works has 4 parts, with core part being compulsory for the course completion and the rest - optional extras, worth 1 credit point each. As figure 1 indicates, the project has several distinct components. The green area highlights the part that is compulsory for the successful completion of the project. It includes the planning of the core network, as well as the access 1
Datacenter Corporate network Core network Cellular access network Access network Access network Figure 1: Project structure network for the users. This will teach you how to apply your knowledge in network dimensioning to create a scalable and efficient network core, as well as cost-effective access network to deliver the triple-play services to the customers. The other parts are as follows: Blue area is the datacenter planning exercise. You will get some experience dimensioning a local data center to be rented out for cloud service proviers such as Google, Netflix and others. Here you will get to play around with some serious bandwidth, as well as serious costs involved in deployment of datacenter solutions. Red area is a cellular network core, which will serve the base stations all over the city, delivering 4G access to the subscribers, as well as some of the beyond-4g services for smart sensors. Unlike conventional access network, cellular coverage is significantly less oriented on traffic and more on reliability and cost-efficiency. Yellow area highlights a corporate network, which presents new challenges in security, performance and availability. Corporate networks require very special topologies, as well as accommodation of some conflicting requests by the picky customers. Please contact the teacher if you want to do any of those! You may also add those to your network later to get more credits. 2 Core network planning Any city-scale network needs a core part and an access network, and this project will teach you how to plan those, such that the costs are minimized and profits are maximized. We will operate under constraints of the communications law of EU, as well as reasonable policy of never lying to our customers. 2
2.1 The situation Somehow, over one night, a city of <city of your choice in Finland that has 100000-300000 people> lost all of its networks. Basically, some highly trained gremlins dug up every single cable and ruined all the data centers. Now that existing operators are bankrupt, you get to take the sweet spot for yourself. You may want to share a city with some of your friends, and you HAVE to share it if you take Tampere as your city of choice. 2.1.1 The demand The users in the city are distributed as shown on the population map. The population map can be freely downloaded from the paikkatietoikkuna.fi web service. It was also made available in qgis format on the course webpage. Obviously, for the purposes of the project we will simplify the map data a little bit. In particular, you will not need to care about things like age groups. The population map aggregates people in polygons that are roughly 1x1 km big. For each of those, you will need the following data: Position on the map (lat,lon) Total population Of course, in reality you would normally go into much deeper details, but these basic parameters should be sufficient for this project. We will assume that each user requires a certain amount of traffic following the on-off model: During work hours (8 hours/day) you may assume that an average user demands 3.0 Mbit/s of traffic with a probability p a = 0.6, and demands no traffic with probability p i = 1 p a. During evening hours (2 hours/day), some people watch internet TV and demand 10.0 Mbit/s of traffic with a probability p a = 0.4. The rest of the day users are generating idle random load of 0.5 Mbit/s of traffic with a probability p a = 0.1. You are free to approximate the actual on-off flow with Bernoulli arrivals or even exponential arrivals if you so wish, it will have minimal impact on your network when 10000 streams are aggregated. You may also assume that traffic is downlink-only and originates from Internet. 2.2 The costs of equipment 2.2.1 Cabling You may assume that within 1 kilometer block ( last mile ) the cables cost 50 per user. You may assume that no modems are needed for users. The last mile cables hold 100 Mbps bitrate so you do not need to worry about their capacity. 3
Root backup link tier 2 concentrator tier 1 concentrator "last mile" points of presense Figure 2: Several tiers of network The cabling for the broadband links costs 1000 per kilometer, and each of the cables supports a bitrate of 10 Gbps. In addition to the fiber itself, the digging adds an extra 5000 per kilometer irrespective of cable capacity. 2.2.2 Switches and routers A switch port for aggregation of data costs 50. The switches perform aggregation only and are incapable of forwarding traffic through. The switches handle up to 10Gbps of traffic on a port. A broadband MPLS router port has a cost of 500, and serves up to 10 Gbps of traffic. You may assume that routers and switches can have arbitrary number of ports. 2.2.3 Other things You will find plenty of lakes and swamps in Finland. Difficult terrain makes deployment of equipment more expensive - add 5000 for putting any of your equipment on water (except for cables). For each data center you also need to pay upkeep money for electricity. That adds up to 5 per 10Gbps link per day. 2.3 Initial planning steps 1. Take the map data and make sure you understand how to manipulate it. Keep in mind that you will need to manipulate it many times, so make sure you save your scripts. 4
2. Use the clustering algorithm to separate your population into K i clusters, and assign an aggregation point to each one of these clusters. You now have your tier 1 routers (see Figure 2 for details). 3. Write up a cost function that will be able to compute the cabling cost based on distance and necessary capacity (a) Keep in mind it must take into account the traffic pattern! 4. Write up another cost function that will be able to figure out the structure and cost of each data center that you will need based on the links that are being aggregated and their capacity (a) Keep in mind it must take into account the traffic pattern, and also produce the distribution for the aggregated flow that you will feed to the next layer 5. You should be able now to easily compute the cost of your last mile and tier 1 data centers. It s time to connect them! 6. Use the positions of tier i 1 data centers as samples for clustering algorithm to figure out the positions of tier i data centers. This time, aim for K i 1 2 K i 1 clusters. (a) To make the clustering behave you may wish to place several samples per data center depending on the average load that it generates. Since your cable cost function is non-linear, it is cheaper to aggregate traffic quickly, as you pay for the digging only once. 7. Once you have just about 3-4 data centers left on the last tier, connect them all to one central point - that would be the root of your network tree, and also the primary gateway. (a) You should still be able to compute the cost of the entire thing easily and automatically, if that is not the case - refactor your code! 3 Optimization steps Now it is time to see if we need to optimize the network! Write a function that computes the aggregate outage probability for your network. The outage probability is defined as the average amount of time the device will be offline during the usage period. Use the following data: Cable: p oc = 1e 7 C, where C is capacity in Gigabits Switch port p os = 1e 6 5
Router port p or = 1e 5 You may also use conventional reservation techniques: N+1, N+2, or even 2N if necessary to improve those basic numbers. Assume that switching in case of failure happens automagically and immediately. Now that you can compute the outage probability you will need to calculate what it is in your network for all of your users. Your optimization constraints are as follows: p o > 1e 3 - you lose your license to operate (and fail the project) For each hour of outage, you pay a fee of 100 per each customer. For each hour of network overload (e.g. when customers are served, but at insufficient rate) you pay a penalty proportional to the outage penalty (e.g. if you promised 10 Mbps and deliver only 1 you pay 100/10 = 10 penalty for each hour). 3.0.1 Optimization advices Your primary optimization techniques will be based around adjusting the clustering parameters K i and locations of the data centers. Keep in mind that the more data a certain center handles the more impact it has on stability and performance of the network To improve reliability of something you can always put it there twice. Or three times. Assume uncorrelated failures on all devices. Most of your problems are non-convex! This section will be updated later 4 Data center planning To be announced 5 Cellular network planning To be announced 6 Corporate network planning To be announced 6