The Effectiveness of Configuration Merging in Point-to-Point Networks for Module-based FPGA Reconfiguration

Communications infrastructure for modular reconfiguration of FPGAs needs to support the changing communications interfaces of a sequence of modules. In order to avoid the overheads incurred by a bus system or network-on-chip, the approach we have taken is to create point-to-point wiring harnesses to support the dynamic intermodule communications. These harnesses are reconfigured at various stages in the application as necessary. The COMMA methodology implements applications on tile-reconfigurable FPGAs such as the Virtex-4. This paper outlines the methodology and describes greedy and dynamic programming-based algorithms for merging configurations, which is a central process in generating wiring harnesses within the methodology. The effects of merging the configuration graphs were explored with both algorithms for a range of device sizes and architectural parameters. Our evaluation indicates graph merging using the greedy method can reduce reconfiguration delay by up to 60% and the dynamic programming algorithm can achieve a further 50% reduction in reconfiguration delay.