Adaptive scheduling framework for multi-core systems based on task-parallel programming model

Due to the poor scalability and intensive competition for computing resource in multi-core runtime system, how to efficiently exploit the advantages offered by multi-core processors has become an imminent challenge. Based on dynamic parallelism feedback strategy, we present an adaptive scheduling framework, named A-SYS, which can dynamically redistribute and manage the computing resources among different applications. A prototype system is implemented based on Cilk and OpenMP runtime systems. The experimental results show that A-SYS improves overall system performance and resource utilization for multi-core systems. Comparing with scheduling algorithms which only rely on runtime systems, A-SYS shortens running time of applications for nearly 50%, and with the number of applications increasing the effect is more significant.