Configuration scheduling using temporal locality and kernel correlation

This paper examines runtime decisions to configure hardware accelerators or execute in software. Traditionally, reconfigurable FPGAs are reconfigured on-demand with the hardware accelerator, as it is needed by the application. If the software kernel which the hardware kernel replaces is available too, then more sophisticated decision making on reconfigurations may lead to improved execution time and reduced power consumption. The temporal locality algorithm is proposed for applications where individual kernels dominate during differing execution modes. The kernel correlation algorithm is proposed for applications where sequences of kernels are invoked in regular patterns. SystemC simulation is used to compare these two scheduling algorithms against the on-demand policy. Both timing and power consumption results are presented. They indicate that a fairly large reconfiguration time is required for configuration scheduling to be beneficial.