GifT:A Gravity-Directed and Life-Time Based Algorithm for Temporal Partitioning of Data Flow Graphs

In reconfigurable systems, reconfiguration latency has a significant impact on the system performance. In this work, a temporal partitioning algorithm is presented to partition data flow graphs for reconfigurable computing systems. Life-time of a node in a data flow graph represents the number of times it executes during the application run. A new factor, called residing eligibility, inspired from the Universal gravitation law, is introduced to depict the eligibility of a node to stay in succeeding partitions and to prohibit it from being swapped in/out. Life-time, size, and distance of a node from the nearest identical node in the data flow graph are important factors for determining its residing eligibility. Assigning nodes to a partition according to their residing eligibility can cause fewer nodes with different functionalities to be assigned to subsequent partitions. Thus, reconfiguration overhead time and also wasted hardware space decreases due to common parts in subsequent configurations on a partially programmable hardware