Optimization of Torus with Segmented Recon gurable Busfor Data

Data routing is one of the most important problems in parallel networks. Long latency in low dimensional networks can be solved by including segmented reconngurable buses (SRB). In addition to physical limitation we choose the segment length such that we may get the best resource utilization and minimize the average latency. After choosing the segment length, we have to arrange these segments carefully to minimize the interaction between them. We deene an alignment metric which can give us the information of traac for a given alignment. We use a prob-abilistic algorithm called hill-climbing to break deadlock. Hill-climbing is shown to be able to eliminate deadlock and stay livelock-free in both theory and simulation with negligible penalty or even better performance.