Context in source publication

Context 1

... lower bound corresponds to a situation when all networks base their decisions only on local knowledge (Eq. 2) and is further referred as local knowledge reference. V. S IMULATIONS The performance and functionality of the algorithms have been evaluated through multiple simulation runs. We have implemented both versions of our algorithm in the OMNet++ environment [26]. In Algorithm A , the network gets the information about how many other networks are requested by the same mobile node. As no other information is available, we assume that the probability of being assigned to any of the networks is equal for all participating networks. In Algorithm B , each network shares its information with exactly one more network. In our testing scenario, these networks do not overlap. We compare the algorithms with the global knowledge reference and the local knowledge reference. For the sake of simplicity, we simulate a scenario with four wireless networks, which covers quite well the scope of the evaluation. In this scenario, a group of users from one network is about to move from one cell of the network to another cell of the same network that experiences a shortage of available bandwidth. In consequence, the cell that the users move to is not able to accommodate all these users. For our evaluation, we run tests with 100, 200, and 300 users moving to this congested cell. Further, we divide the users into four categories in terms of requested bandwidth. To define these categories we use service class characteristics defined by Tragos et al. [20] as follows: a ) at 64 kbps, for simple telephony and messaging b ) at 512 kbps, for web browsing c ) at 1024 kbps, for interactive media and d ) at 2000 kbps, for video streaming, each category having approximately the same number of users. Note that none of the networks have enough resources to accommodate all users alone. All four networks must be used in order to meet the requirements of all users. We run also tests when total bandwidth of all networks is not sufficient to accommodate all users. The tests are done for network conditions that result in 5%, 10%, 15%, 20%, 25%, 30% dropped calls if the global knowledge reference is applied. The time τ i,j for the user j to stay in the network n i before performing a horizontal handoff, or a cell residence time, is randomly distributed in the range [1 , 100] time units. We evaluate how mobile nodes are distributed among the networks after one iteration of the algorithm run. We calculate the number of decision errors as a number of users whose connection ends up in dropped calls due to wrong network allocation. These errors are the results of wrong assignments to networks that do not have sufficient bandwidth to accommodate the assigned users. For each group of users (100, 200, 300 users), we repeat the experiment 1000 times with different sets of τ i,j . For all tests done, the top and bottom 5 % of the results are excluded from the evaluation. The results are averaged over these simulation runs and are depicted for minimum value results in Figure 4(a), for average value results in Figure 4(b), for maximum value results in Figure 4(c), for cumulative distribution function in Figure 5. The values 0 in the results for Algorithms A and B in Figure 4(a) indicate that all users are assigned to the networks correctly. The global knowledge reference is 0 for all experiments meaning that in the centralized solution, all users were assigned to the networks without any dropped calls. The results with dropped calls in the global knowledge reference are depicted in Figure 6. The figure shows the results for 200 mobile nodes. The results for 100 and 300 mobile nodes are very similar to the results for 200 nodes and therefore are not included in the paper. The tests show that both Algorithms A and B can distribute the users between the networks significantly better than the local knowledge reference. Algorithm A performs better than Algorithm B for all user groups through all tested values for dropped calls in the optimum solution. It shows that sharing partial information in Algorithm B makes little use of extra information from just one network. However, Algorithm B requires significantly more information to exchange between the networks than Algorithm A . It also requires more sophisticated mechanisms and protocols to be implemented in the networks, including security considerations and synchronization of the information flow. We also evaluate the dynamic scenario. For these tests, the algorithms are run until all clients are assigned to the networks with sufficient bandwidth, also considering the arriving calls. The arrival rate of new calls is modeled with a Poisson stream. The graphs depicted in Figure 7 show the averaged results for 100, 200, and 300 users over 1000 test runs. The x-axis shows the number of iterations of the algorithm. The y-axis shows the percentage of decision errors. Clearly, both Algorithms A and B converge faster than the local knowledge reference. There is very little difference between Algorithms A and B even though Algorithm B relies on more information. We estimate signaling overhead S o for the algorithms and the references. As signaling required to trigger network selection is the same for the references and the algorithms these messages are excluded from the estimation. For the global knowledge reference all n networks in consideration need to exchange the information about m users that get triggered network selection, and the overhead is estimated as follows (Eq. ...