| The goal of this thesis is to study and improve the benefits of using distributed computing for solving power system applications, in particular transient stability analysis (TSA), which is one of the computationally most demanding power system applications.;In distributed computing, several important factors, like parallelism, load imbalance and communication overhead need to be considered. In order to be able to study these factors, we implement three different algorithms for TSA on a network of workstations and on SP2 systems. One method is a parallelization of the fastest sequential algorithm known. The second method is the W-matrix method, which improves parallelism but has high communication requirements. The third method increases parallelism at the expense of additional computations. Results from running experiments using real power system data allow us to study effects of parallelism, load imbalance and communication overhead on the performance of the algorithms. We support the experimental findings by means of theoretical performance and scalability analysis. Our theoretical analysis clarifies experimental results and allows performance predictions for larger systems being simulated or larger number of processors being used.;From experimental and theoretical results it becomes clear that the dominant performance-limiting factor is communication performance. We alleviate the effect of this bottleneck for SP2 systems by exploiting the knowledge of the underlying network when suggesting efficient implementations of frequently used communication patterns. First, we present algorithms which route general permutations using centralized control and linear-complement (LC) permutations using distributed control in one pass through the SP2 switches. Then, we show how collective communications, the most frequently used communication patterns in many applications, can be expressed as sequences of partial LC permutations and be routed in the minimum number of communication steps through SP2 switches without contention. Our implementations of collective communications guarantee that no contention will occur and that collective communications will be routed in the minimum number of communication steps. Consequently, they will considerably improve the performance of our implementations as well as many other applications running on SP2 systems. |
