Using dynamic partial reconfiguration of FPGAs in real-Time systems

摘要

The use of hardware accelerators to implement computationally intensive tasks in real-time systems can lead to a reduction of the worst-case execution time (WCET). An additional potential benefit is that a WCET-analysis may be simpler to perform because hardware generally has a more time-predictable behavior than software. The dynamic partial reconfiguration (DPR) feature offered by modem FPGAs allows accelerators that are no longer needed to be replaced with new ones, leading to more efficient utilization of hardware resources. This paper presents an experimental evaluation of the potential benefits of using DPR to implement hardware accelerators in real-time systems, focusing on trade-offs between hardware utilization, worst-case performance, and speed-up over a pure software solution. Moreover, it also investigates the trade-off between the use of multiple specialized accelerators combined with DPR instead of the use of a more general accelerator, and the memory footprint of the partial-bit streams. The experiments show that DPR in combination with accelerators results in: (i) better utilization of the FPGA resources, (ii) performance that is comparable with non-reconfigurable solutions, and (iii) tighter WCET bounds.