An approximation algorithm for scheduling on heterogeneous reconfigurable resources

Abstract

Dynamic reconfiguration imposes significant penalties in terms of performance and energy. Scheduling the execution of tasks on a dynamically reconfigurable device is therefore of critical importance. Likewise, other application domains have cost models that are effectively the same as dynamic reconfiguration; examples include: data transmission across multiprocessor systems; dynamic code updating and reprogramming of motes in sensor networks; and module allocation, wherein the sharing of resources effectively eliminates inherent reconfiguration costs. This article contributes a fully polynomial time approximation algorithm for the problem of scheduling independent tasks onto a fixed number of heterogeneous reconfigurable resources, where each task has a different hardware and software latency on each device; the reconfiguration latencies can also vary between resources. A general-purpose processor and a field programmable gate array were used to experimentally validate the proposed technique using a pair of encryption algorithms. The latencies of the schedules obtained by the approximation scheme were at most 1.1× longer than the optimal solution, which was found using integer linear programming; this result is better than the theoretical worst-case guarantee of the approximation algorithm, which was 1.999×. The length of the schedules obtained using list scheduling, a well-known polynomial time heuristic, were at most 2.6× longer than optimal.