Off-line Load Balancing

We implemented an external feedback for the load balancing. During run time we collect the execution time of each link and each intersection (node). The statistics are output to file every 1000 time-steps. For the next run of the micro-simulation, the file is fed back to the initial load balancing algorithm. In this iteration, instead of using the link lengths as load estimate, the actual execution times are used as distribution criterion. Fig. 10 shows the new tiles after such a feedback (compare to Fig. 6).

To verify the impact of this approach we monitored the execution times per time-step throughout the simulation period. Figure 11 depicts the results of one of the iteration series. For iteration 1, the load balancer uses the link lengths as criterion. The execution times are low until congestion appears around 7:30 am. Then, the execution times increase fivefold from 0.04 sec to 0.2 sec. In iteration 2 the execution times are almost independent of the simulation time. Note that due to the equilibration, the execution times for early simulation hours increase from 0.04 sec to 0.06 sec, but this effect is more than compensated later on.

The figure also contains plots for later iterations (11, 15, 20, and 40). The improvement of execution times is mainly due to the route adaptation process: congestion is reduced and the average vehicle density is lower.

  

Figure 10: Partitioning after load balancing. Compare to Fig. 6.

  

Figure 11: Execution times with external load feedback