Multi-objective exploitation of pipeline parallelism using clustering, replication and duplication in embedded multi-core systems

With the popularity of mobile device, people require more computing power to run emerging applications. However, the increase in power consumption is a major problem because power is quite limited in embedded systems. Our goal is to consider power consumption along with latency and throughput. We proposed a heuristic algorithm, called Parallel Pipeline Latency Optimization for high performance embedded systems (PaPiLO), based on clustering, replication and duplication, to minimize latency under power and throughput constraints. Experimental results show our method can get 15% latency reduction and 10% improvements for random task graphs and MPEG-2 decoder, respectively.