DVS for buffer-constrained architectures with predictable QoS-energy tradeoffs

We present a new scheme for dynamic voltage and frequency scaling (DVS) for processing multimedia streams on architectures with restricted buffer sizes. The main advantage of our scheme over previously published DVS schemes is its ability to provide hard QoS guarantees while still achieving considerable energy savings. Our scheme can handle workloads characterized by both, the data-dependent variability in the execution time of multimedia tasks and the burstiness in the on-chip traffic arising out of multimedia processing. Many previous DVS algorithms capable of handling such workloads rely on control-theoretic feedback mechanisms or prediction schemes based on probabilistic techniques. Usually it is difficult to provide QoS guarantees with such schemes. In contrast, our scheme relies on worst-case interval-based characterization of the workload. The main novelty of our scheme is a combination of offline analysis and runtime monitoring to obtain worst case bounds on the workload and then improving these bounds at runtime. Our scheme is fully scalable and has a bounded application-independent runtime overhead.