next up previous contents
Next: 6.2.2 Robustness Up: 6.2 Distributed Computing Previous: 6.2 Distributed Computing

6.2.1 Workload Distribution

Although there is no need for information exchange between concurrent parts of the parallel simulations, the completion of the whole simulation depends on the very last part to finish. This means that load balancing is necessary in order to reduce the overall simulation time. Load balancing means that faster machines should get a bigger share of the whole workload and, on the other hand, slower machines should obtain less of it. Thus, since the faster machines do more work than the slower ones do, the former will have to wait less until the latter finish. The ultimate case occurs when workload is optimally balanced and all parts finish simultaneously. A strategy which intends to achieve this situation is discussed in Section 6.2.4.

Rudi Strasser