XIE ET AL.:ASSOCIATION CONTROL FOR VEHICULAR WIFI ACCESS:PURSUING EFFICIENCY AND FAIRNESS 1329 TABLE 1 TABLE 2 Parameters for Generating Vehicle Traffics Abbreviations parameters default value parameters default value Abbreviation Full Name accel 0.8(m/s2) length 5(m) SSF Strongest Signal First decel 4.5(m/s2) speed 15(m/s) CUB Connect Until Broken sigma 0.5 depart 0(s) OPT-E(offline Offline optimization for efficiency period 30(s) repno 100 OPT-E(online) Online optimization for efficiency density 22 (users/km) OPT-PF(offline) Offline optimization for proportional fairness OPT-PF(online) Online optimization for proportional fairness OPT-MM Online optimization for max-min fairness (5 m/s,20 m/s).We simulate 500 various routes over the road network for the vehicles,for each route we randomly pick the origin position and the destination position and at any location of the roads.In the following part,we hence specify the trip for the route.We use depart to denote conduct the performance evaluation to show how the the time at which the vehicles are emitted into the network, optimal solutions work under the two different situations. and we use period to denote the average time interval after To obtain each simulation result,we take the average value which another vehicular user with the same route shall be of 50 simulation runs.In the supplementary material,which emitted and repno to denote the number of vehicles to emit can be found on the Computer Society Digital Library at which share the same route.We use random seeds to http://doi.ieeecomputersociety.org/10.1109/TPDS.2011.17, generate the time intervals between emitted vehicles with we illustrate more detail simulation results in a compre- the same route.As we set period=30s,thus on average hensive approach. every 30s the vehicles are emitted,and we set the 95 percent 5.1 Efficiency and Fairness confidence interval of the time interval as (20 s,40 s). In this section,we evaluate the performance in terms of Therefore,the application scenario involves about 50,000 efficiency and fairness.In order to illustrate the perfor- vehicular users and lasts about 50 minutes.According to the mance gains of our optimized solutions,we compare with above settings,the average density density 22 users/km, two heuristic strategies.The first strategy is Strongest Signal i.e.,the average number of vehicles per kilometer of road is First,which always associates a user with the Ap yielding 22.We observe that the 95 percent confidence interval for the the strongest received signal strength at all times.The density is(14 users/km,60 users/km). second strategy is Connect Until Broken,which maintains a We conduct the performance evaluation based on the connection with a user and an AP until the user considers settings of large road network and vehicle traffic generated the link to be broken.Upon disconnection,the user will be by SUMO.We randomly place the APs inside the specified associated with a new AP which yields the largest signal region and adopt the experiment results from [1]to strength.When calculating the optimized solution,we solve simulate the effective bit rates of APs.We set their peak the linear program and convex program using MATLAB.In bit rates within the range from 4,000 to 5,000 kbps for the rest of this paper,we use the abbreviations as shown in vehicular users.To sufficiently evaluate the performance of Table 2 to denote the specified solutions.For the ease of various association control strategies,we consider two comparison,we set w;=1 for each user.We set e=1 kbits kinds of situations for AP deployment:the dense AP for OPT-PF(online),and,respectively,set C=200 kbps and deployment and the sparse AP deployment.For the dense C=0 kbps in the dense AP situation and sparse AP AP deployment,we randomly deploy 500 APs and make situation for both OPT-E(offline)and OPT-E(online). sure that at any location of the roads the user is within Fig.4a depicts the total throughput of all users achieved effective range of at least one AP.For the sparse AP by various solutions.As in each run of simulation the deployment,we randomly deploy 150 APs and the user is generated traffic mobility has some variances,in order to not guaranteed within the effective range of at least one AP show the statistical performance results,so we provide the 1200 ■SgF CUB OPT-E(omine) 1000 -CPT-Eonine) OeT-PEIcttlne OPT-E(online) CPT-PF[online CPT-PF(omline 800 -OPT-MM CPT-PFlonline OPT-MM 600 400 200 (b)Sparse AP conditiong 0.5 15 *10 (a) (b) Fig.4.Simulation results for efficiency and fairness.(a)Total throughput(kbps)for all users.(b)Per-user throughput comparison with dense AP deployment.ð5 m=s; 20 m=sÞ. We simulate 500 various routes over the road network for the vehicles, for each route we randomly pick the origin position and the destination position and hence specify the trip for the route. We use depart to denote the time at which the vehicles are emitted into the network, and we use period to denote the average time interval after which another vehicular user with the same route shall be emitted and repno to denote the number of vehicles to emit which share the same route. We use random seeds to generate the time intervals between emitted vehicles with the same route. As we set period ¼ 30 s, thus on average every 30 s the vehicles are emitted, and we set the 95 percent confidence interval of the time interval as ð20 s; 40 sÞ. Therefore, the application scenario involves about 50,000 vehicular users and lasts about 50 minutes. According to the above settings, the average density density ¼ 22 users/km, i.e., the average number of vehicles per kilometer of road is 22. We observe that the 95 percent confidence interval for the density is (14 users/km, 60 users/km). We conduct the performance evaluation based on the settings of large road network and vehicle traffic generated by SUMO. We randomly place the APs inside the specified region and adopt the experiment results from [1] to simulate the effective bit rates of APs. We set their peak bit rates within the range from 4,000 to 5,000 kbps for vehicular users. To sufficiently evaluate the performance of various association control strategies, we consider two kinds of situations for AP deployment: the dense AP deployment and the sparse AP deployment. For the dense AP deployment, we randomly deploy 500 APs and make sure that at any location of the roads the user is within effective range of at least one AP. For the sparse AP deployment, we randomly deploy 150 APs and the user is not guaranteed within the effective range of at least one AP at any location of the roads. In the following part, we conduct the performance evaluation to show how the optimal solutions work under the two different situations. To obtain each simulation result, we take the average value of 50 simulation runs. In the supplementary material, which can be found on the Computer Society Digital Library at http://doi.ieeecomputersociety.org/10.1109/TPDS.2011.17, we illustrate more detail simulation results in a compre￾hensive approach. 5.1 Efficiency and Fairness In this section, we evaluate the performance in terms of efficiency and fairness. In order to illustrate the perfor￾mance gains of our optimized solutions, we compare with two heuristic strategies. The first strategy is Strongest Signal First, which always associates a user with the AP yielding the strongest received signal strength at all times. The second strategy is Connect Until Broken, which maintains a connection with a user and an AP until the user considers the link to be broken. Upon disconnection, the user will be associated with a new AP which yields the largest signal strength. When calculating the optimized solution, we solve the linear program and convex program using MATLAB. In the rest of this paper, we use the abbreviations as shown in Table 2 to denote the specified solutions. For the ease of comparison, we set wj ¼ 1 for each user. We set ¼ 1 kbits for OPT-PF(online), and, respectively, set C ¼ 200 kbps and C ¼ 0 kbps in the dense AP situation and sparse AP situation for both OPT-E(offline) and OPT-E(online). Fig. 4a depicts the total throughput of all users achieved by various solutions. As in each run of simulation the generated traffic mobility has some variances, in order to show the statistical performance results, so we provide the XIE ET AL.: ASSOCIATION CONTROL FOR VEHICULAR WIFI ACCESS: PURSUING EFFICIENCY AND FAIRNESS 1329 TABLE 1 Parameters for Generating Vehicle Traffics TABLE 2 Abbreviations Fig. 4. Simulation results for efficiency and fairness. (a) Total throughput (kbps) for all users. (b) Per-user throughput comparison with dense AP deployment.