Cost-optimal VPN based VoIP network design
Transcription
Cost-optimal VPN based VoIP network design
Cost-optimal VPN based VoIP network design * Dániel Orincsay, Balázs Gábor Józsa, Levente Tamási Ericsson Research Hungary, Traffic Analysis and Network Performance Laboratory, Budapest University of Technology and Economics, Dept. of Telecommunications and Media Informatics E-mail: {Daniel.Orincsay, Balazs.Jozsa, Levente.Tamasi}@eth.ericsson.se Abstract This paper addresses the design of cost-optimal VoIP networks. In our model the whole VoIP network is divided into two logical components: the access network and the transport network. The access network consists of the VoIP end-points that connect to the transport network through an edge router. Since more edge routers can be available for a given VoIP node, one task of the design is to assign an edge router for every VoIP node. The edge routers serve as gateways for the assigned VoIP nodes towards the transport network. The gateways have to be connected in such way that security and availability could be assured for the VoIP traffic. One obvious approach to realize this is to define a VPN. Supposing large volume of VoIP traffic the cost of the VPN can be significant. Thus the other task of VoIP network design is to specify the transport VPN in the most economical way. These two tasks of VoIP network design can be solved separately using existing methods. Nevertheless, the VoIP region specification influences the cost of the final solution to a great extent. Therefore, in this paper we propose a novel solution, in which the edge router assignment takes the objective function of VPN specification into consideration. Moreover, a numerical analysis is provided with the help of simulations. Keywords: VoIP, VPN, cost-optimal, network design 1 Introduction Nowadays, the all-IP concept is favored by the infocommunication industry, intending to conduct all different types of traffic over the Internet protocol dominant in the networking area. As part of all-IP an increasing number of companies in the telephony area commit themselves to using the voice over IP (VoIP) technology. Assuming a large VoIP network having a huge amount of customers it is necessary to take QoS as well as economical goals into account during the design phase. In our model the VoIP network consists of two logical components: the access network, and the transport network. The access network includes the VoIP nodes, namely, the customer end-points intending to use the VoIP service. The transport network serves for carrying the aggregated VoIP traffic between the various access areas. Its main parts are the edge routers, the transit routers, and the physical links between them. A VoIP node can reach more than one edge router fulfilling the QoS requirements (e.g., limited maximal delay) of the phone call service. Therefore, it is necessary to select one of them that will serve as gateway towards the transport network. The VoIP nodes assigned to the same edge router form a so-called VoIP region. * Generally, it is more economical to apply a virtual private network (VPN) instead of deploying a brandnew physical network to realize the transport network for a VoIP service (see e.g., [1], [2]). Thus we follow this approach in this study. The cost of the VPN depends on the capacity of corresponding devices, such as routers and links. The area of cost-optimal VPN design has been widely studied in the literature, where beside the topology the set of traffic demands and the cost functions of devices were supposed as input parameter. However, during VoIP network design the traffic distribution between the VPN nodes cannot be considered as a fix input, since it largely depends on the specification of VoIP regions. In summary, two main tasks can be differentiated in the case of VoIP network design: (1) VoIP region specification, i.e., the assignment of each VoIP node to exactly one VPN edge router, and (2) the design of the transport network covering the selected VPN nodes. These two tasks can be solved independently by applying existing methods, which means that the objective of VPN transport network design is disregarded during the VoIP region specification. However, it can be more efficient to take also cost and quality factors concerning the transport network to be composed into consideration already in the first task. This issue is addressed in this paper by making sev- This work was supported by the Ministry of Education, Hungary, under the reference No. IKTA-0092/2002. http://w3.ttt.bme.hu/ikta-2002 eral propositions to solve the VoIP region specification task of the VoIP network design problem. In order to solve the VPN transport network design after specifying the VoIP regions a core network design algorithm (CND) is used, which proved to be efficient in the field of cost-optimal VPN design [3]. The numerical investigation of proposed methods is based on the result of CND—performed after each region specification method—that gives the final cost of the VoIP network. The rest of the paper is organized as follows. The next section describes the used network, traffic, and cost models. It also includes the formulation of the VoIP network design problem. In Section 3 novel approaches are introduced aiming at solving the VoIP region specification. Then Section 4 presents numerical results obtained from the performed simulations. Finally, the conclusions are drawn. 2 Problem statement This section introduces the interpretation of VoIP network design problem considered in the paper. First, the applied network, traffic, and cost models are discussed. Then the formulation of the problem follows. Finally, the optimization objective is presented. 2.1 Network model The VoIP network is modeled by a graph in the following way. Each customer end-point that uses the VoIP service is called VoIP node. The set of VoIP nodes is denoted by W. The possible edge routers of the VPN transport network to be composed are called VPN nodes. The set of VPN nodes is denoted by V. Further, we are given so-called VoIP edges connecting the VoIP-VoIP and VoIP-VPN node-pairs. Each VoIP edge f has a delay attribute delayf assigned representing its maximal one-way latency. Based on these edge delay values the delay value dwv can be determined for each VoIP node w∈W and VPN node v∈V pair, representing the minimal guaranteed latency between them. If a VPN node v is not available for VoIP node w, value of dwv is considered as ∞. In the current interpretation the QoS requirement—that the route between the VoIP node and the corresponding VPN gateway has to fulfill—refers to the maximal access network latency dmax. Physical connections between the VPN nodes are called VPN edges, and their set is denoted by E. The use of VPN nodes as well as VPN edges is optional, generally only a subset of them is included in the final solution of design. The bandwidth of VPN nodes and edges is limited, their particular capacity values are determined during the design using the corresponding cost functions (see Section 2.3). 2.2 Traffic model Although a VoIP node may refer to one particular customer owning a VoIP phone, it generally represents a private branch exchange (PBX) including a VoIP media gateway. Due to the large number of VoIP nodes the generally applied approach of the pipe model [4] (also known as trunk model)—that means the sourcedestination pair based handling of traffic—is inefficient. Therefore, the hose model [5] is followed, which defines only the sum incoming and outgoing traffic of a node. Considering that telephony calls are handled, the incoming traffic and outgoing traffic are equal referring to the symmetric hose model. Thus, the traffic of VoIP node w is modeled by a bandwidth demand value trw that shows the amount of capacity needed to satisfy the calls generated (and received) by the VoIP users in the given node. This value can be derived from the number and calling behaviors of different VoIP users. However, this issue is related to the area of traffic modeling, and it is out of the scope of the paper. Although the hose model based design results in such networks that can accommodate extreme traffic distributions as well, the high amount of spare capacity and consequently extra price make it unacceptable in cost-sensitive situations. Therefore, in the case of the VPN transport network the pipe model is estimated in the following way. First, the hose traffic values of VoIP nodes in the same region are summed representing the total hose traffic of the corresponding VPN edge node. Considering that these traffic values represent large number of users and the number of VPN nodes is relatively low, the pipe model can be approximated well in the following way. The summed traffic of a particular VPN node is distributed among the other VPN nodes in direct proportion to their total hose traffic. 2.3 Cost model In order to model the cost/capacity dependencies of devices realistically stepwise cost functions are used. Although this approach makes the design problem mathematically complex, it can fulfill the high accuracy requirements arising in real situations. Moreover, in our model each VPN node v and VPN edge e has individual cost function costv and coste, respectively, which enables to take special cost modifying factors as well as policy reasons into account. 2.4 Problem formulation The VoIP network design problem has the following input parameters: (1) the set of VoIP nodes W, (2) the set of possible VPN nodes V, (3) the set of possible VPN edges E, and (4) the set of delay values dwv for all VoIP node w∈W and VPN node v∈V pairs (derived from the structure of VoIP nodes and edges). As the output of the region specification, each VoIP node is assigned to exactly one VPN node, which determines a set of VPN nodes that are mandatory elements of the VPN transport network to be formed. Further, as the result of the design process, the VPN transport network is specified including the exact capacity values of all devices. The paths to be established corresponding to the aggregated VoIP traffic demands are also given by the applied design method. 2.5 minimize ∑ coste (load e ) + ∑ costv (load v ), (1) v∈V e∈E where load refers to the actual capacity need on a certain device, and cost(load) indicates the corresponding price of this device. Methods This section proposes several methods that are capable of solving the VoIP region specification problem. Since it is related to the so-called set-covering problem widely studied in the literature [6], [7], term ‘covering’ is used throughout the descriptions of various algorithms referring to the situation when a VPN node is selected as gateway for a VoIP node. 3.1 Greedy covering algorithm (GC) This section presents greedy algorithms that are based on a well-known greedy solution of the set covering problem (see e.g., [8]). As the first step such VoIP nodes are searched for that reach only one VPN node fulfilling the delay requirements. This defines a set of VPN nodes that is a mandatory part of the solution. Consequently, we consider covered all VoIP nodes that can be covered by these VPN nodes. The second step is an iteration in which further VPN nodes are selected one by one based on a certain utility value. When selecting a particular VPN node the corresponding (still uncovered) VoIP nodes get covered. The iteration stops if the selected VPN nodes cover all VoIP nodes. In the following, two utility value variants are presented. 3.1.1 3.1.2 VPN node cost based GC (GC-C) The utility value used by this variant is the number of still uncovered VoIP nodes that the given VPN node could cover divided by the cost of the first capacity step of its cost function. The idea behind this complex utility value is to take also the establishment costs of the VPN nodes into account during the iteration. Objectives Aiming at minimizing the overall establishment cost of the VoIP network, which basically depends on the cost of the VPN transport network, the objective of the whole design process is to: 3 ered by the given VPN node. Thus this approach aims at minimizing the number of selected VPN nodes, however, it does not take any cost factors into consideration. VPN node number based GC (GC-N) In this variant the utility value corresponds to the number of still uncovered VoIP nodes that can be cov- 3.2 Evolutionary covering algorithm (EC) A common property of the above presented greedy covering algorithms is that they assign a VPN node for each VoIP node only once, which means that they do not vary the existing assignments. This approach is called construction method, i.e., the algorithm stops when the first feasible solution is found. This provides fast region specification, however, its drawback is that there is no possibility for sophisticated optimization. This section proposes such algorithms that are based on the well-known paradigm of evolutionary algorithms (also called genetic algorithm) [9], which enables to select among more feasible solutions using complex entity cost calculation (see Sections 3.2.13.2.3). The representation of the VoIP region specification problem applied in the evolutionary algorithm is the following. Each gene of an entity corresponds to a VoIP node, and its value refers to the VPN node the VoIP node is assigned to. During the mutation one gene of an entity is changed randomly, which means that the given VoIP node is re-assigned to another VPN node. During the crossover of two entities, the child inherits its genes from one of its parents with equal probability. The selection of entities that do not survive is done by killing the oldest entity from k randomly selected entities, and killing the worst one (based on the used entity cost calculation) from another k randomly selected entities. In order to create an appropriate initial population the GC-C was applied in the following way. A feasible solution is sought by GC-C, excluding exactly one VPN node from the solution. The excluded VPN node is altered, thus more different feasible coverings consisting of VPN nodes having low establishment cost values form the initial population. The algorithm stops if the cheapest solution—considering the actual cost calculation—did not change during the last n steps. The applied entity cost calculation method has a large influence on the quality of solution. Thus more approaches are investigated in the study, as it can be seen in the following sections. 3.2.1 Cost approximation based EC (EC-C) The main idea behind the cost approximation based method (EC-C) is that we try to foresee the final cost of the VPN transport network to be designed. It routes the set of the aggregated VoIP traffic demands several times, based on different random orders. Dijkstra’s shortest path algorithm is applied for this purpose in which such edge weight function is considered that prefers devices having low cost per unit traffic. Finally, the price of the cheapest configuration is considered as the cost of entity. 3.2.2 Distance weighted traffic based EC (EC-D) When applying the distance weighted traffic based method (EC-D), the product of the bandwidth requirement and the length of the possible shortest path (in terms of hop-count) is calculated for each aggregated VoIP traffic demand. Then the cost of entity is specified as the sum of these products regarding the whole network. 3.2.3 Two-level cost metric based variants (EC-C2, EC-D2) We examine also the two-level variants of the above cost metrics in the following way. The number of selected VPN edge routers serves as primary metric. Then the cost approximation or the distance weighted traffic methods scaled by an importance factor s are regarded as second metric. 4 Results In order to investigate the performance of the proposed algorithms we carried out simulations using artificial problem instances. First, the automated method of problem instance generation is described, then the performed simulation scenarios are presented including the analysis of the numerical results. 4.1 Problem instance generation During the simulations we aimed at creating such problem instances that represent well the real situations. However, we note that in case of other types of problem instances the results of algorithms may differ. The first task is to generate the topology of network including VoIP nodes and edges as well as possible VPN nodes and edges. For this purpose we applied a random VoIP graph generator method that is based on the Barabási-Albert model [10], [11]. This approach is based on the power laws of Internet topology [12], [13], and nowadays it is frequently used to model wide area communication networks. Various size topologies were examined, however, we present results for networks having 500 VoIP nodes, and 50 Fig. 1 Example network topology possible VPN nodes. The delay value of VoIP edges were specified randomly using such distribution that generates values proportionally to the lengths of edges, i.e., the distances between their end-points. The traffic demands of VoIP nodes were generated randomly in the following way. First, the number of maximal parallel calls was generated for each VoIP node in the interval [1, 2n]. The value of n refers to the average number of maximal parallel calls, and it was shifted from 96 to 256. Then the number of parallel calls had to be scaled due to the actually used codec type and packetization overhead (RTP/UDP/IP header). Assuming the use of ITU-T recommendation G.711 PCM codec—that can provide the highest voice quality—with silence suppression having activity factor of 55% and 5 ms packetization time, the investigated VoIP traffic (trw) interval was 6 to 16 Mbps. The cost functions of VPN edges were based on the STM standards, while in the case of VPN nodes three different size devices were assumed. The detailed description of the used cost functions can be found in [1]. 4.2 Optimization of parameters The first simulation scenario targeted the optimization of parameters of the evolutionary algorithm EC. It turned out that parameter value combination k=3, n=1000, s=500 can give a good tradeoff between the quality and running time measures, therefore this parameter value combination was used throughout the next simulation scenarios. 4.3 Total network cost The most important performance indicator is the resulting total cost of the network. As it can be seen in Fig. 2, the simplest method GC-N provided the most expensive solution. Taking the establishment cost of VPN nodes into account (GC-C), about 3% improve- Total network cost [unit] 2000 4.5 GC- N GC- C 1800 EC-C 1600 EC-D EC-C2 EC-D2 1400 1200 1000 6 8 10 12 14 16 Average VoIP network traffic [Mbps] Fig. 2 Total network cost using different algorithms ment can be achieved. Algorithms EC-C and EC-C2 performed similarly, and they decreased the total network cost by further 7–9% compared to GC-C. Applying EC-D more economical configurations can be obtained in the higher traffic intervals. EC-D2 proved to be the best algorithm as it outperformed all other approaches in all examined test cases. The improvement compared to the simplest method GC-N varied from 15–25% depending on the traffic volume. Although in the case of off-line network design the running time has only secondary importance, it is worth examining also this factor in order to make the investigations complete. As it can be seen in Table 1, GC algorithms were very fast as they provided results within one minute on average. EC-C and EC-C2 had similar results again, but one order of magnitude higher compared to GC. Basically, the time consumption of region specification is low in the case of distance weighted traffic based algorithms EC-D and EC-D2. However, EC-D2 resulted in higher overall running time due to the fact that the running time of the final VPN design step carried out by CND highly depends on the number of VPN gateways. Design algorithm Running time [min] Table 1 4.4 Number of VPN nodes Besides the cost of the VPN transport network its size is also an important attribute, which can be described well by the number of VPN nodes. Thus, in this scenario we compared the algorithms focusing on this basic measure. Fig. 3 presents the number of used VPN nodes in the final transport network differentiating the edge and transit types. Since these measures did not show relevant change against the change of average VoIP node traffic, results are presented only for the 12 Mbps value. We can see that as the total network cost decreases, the total number of VPN nodes decreases. The exception is EC-D, whose result differs relevantly from the other ones as it uses much more VPN gateways. The reason for this could be the lower impact of VPN node number on the entity cost values compared to the other cases. Another observation is that the number of VPN edge nodes shows increasing tendency. This means that the sophisticated selection of VoIP traffic aggregation points is more important than keeping their number as low as possible. Number of VPN nodes 50 ed g e t rans it 40 Running time 5 GC-N GC-C EC-C EC-C2 EC-D EC-D2 1 1 12 11 19 6 Running time values Conclusion This paper addressed the area of cost-optimal VoIP network design. The whole design problem consists of two main tasks: the specification of VoIP regions, and the design of the VPN transport network. We have proposed a novel approach aiming at improving the cost efficacy by taking the objective of transport network design into consideration during the region specification. We have presented various new algorithms that realize the novel approach. In order to evaluate their performance, numerous simulations have been carried out. It has turned out that significant reduction in total network cost can be achieved by applying sophisticated cost evaluation in the VoIP region specification phase. Based on the performed simulations the evolutionary algorithm—using VPN node number and distance weighted traffic based cost approximation— seems to be the best choice. Our future work includes the investigation of such situations where also other types of traffic demands having high bandwidth requirement—such as video telephony—are handled. Another direction is to take also reliability issues into consideration during the VPN transport network design. 30 Acknowledgment 20 10 0 GC-N Fig. 3 GC-C EC-C EC-C2 Design algorithm EC-D EC-D2 Number of edge and transit VPN nodes The authors would like to thank Gábor Magyar for his insightful suggestions and comments. Further, the authors are grateful to the members of project IKTA and people in Ericsson Research Hungary for their valuable remarks. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] Krapf, E.: Avaya's VoIP over VPN, Business Communications Review, pp. 62–63, March 2002 VocalTec Communications Ltd., http://www.vocaltec.com Józsa, B.G.; Orincsay, D.; Kern, A.: On the Use of Routing Optimization for Virtual Private Network Design, In proc. of ONDM’2003, Budapest, Hungary, pp. 865–880, February 3–5, 2003 Cisco Systems, Inc.: Cisco IOS MPLS Quality of Service, White Paper, 2003 Kumar, A.; Rastogi, R.; Silberschatz, A.; Yener, B.: Algorithms for Provisioning Virtual Private Networks in the Hose Model, SIGCOMM’01, San Diego, USA, August 27–31, 2001 Caprara, A.; Fischetti, M.; Toth, P.: Algorithms for the Set Covering Problem, Annals of Operations Research, Vol. 98, pp. 353–371, 2000 Beasley, J.; Chu, P,: A genetic algorithm for the set covering problem, European Journal of Operational Research, Vol. 94, pp. 392–404, 1996 Daskin, M. S.: Network and Discrete Location, Wiley, New York, 1995 Bäck, T.; Fogel, D.B.; Michalewicz, Z.: Handbook of Evolutionary Computation, University Oxford Press, New York, 1996 Barabási, A.; Albert, R.: Emergence of scaling in random networks, Science, pp. 509–512, October 1999 Albert, R.; Barabási, A.: Statistical Mechanics of Complex Networks, Review of Modern Physics, Vol. 74, pp. 47–97, January 2002 Faloutsos, M.; Faloutsos, P.; Faloutsos, C.: On Power-Law Relationships of the Internet Topology, In proc. of ACM SIGCOMM’99, August 1999 Bu, T.; Towsley, D.: On Distinguishing between Internet Power Law Topology Generators, In proc. of INFOCOM’2002, June 2002