SAWS:Selective Asymmetry-aware Work-stealing For Asymmetric Multi-core Architectures

Multi-Core architectures have been applied widely in industry,especially to personal computers,smart phones and most fields demanding of high computing performance.While with the dramatical development of computing performance in Multi-Core architectures,more and more drawbacks are to emerge,like power consumption,researchers propose a new architecture ——Asymmetric Multi-Core(AMC)architectures,where cores in different CPUs have different performance and power consumption.And when multiple CPUs are integrated in the same computer,the memory is often organized as Non-Uniform Memory Access(NUMA)structure.AMC architectures with NUMA-based memory systems have been used from large-scale datacenters to mobile smart-phones for their high performance as well as energy efficiency.However,existing task scheduling policies often result in the poor performance of parallel programs on emerging AMC architectures due to the unbalanced workload,the severe shared cache misses and remote memory accesses between NUMA memory nodes.To solve this problem,we propose a Selective Asymmetry-aware Work-Stealing(SAWS)runtime system,which can reduce remote memory accesses while balancing workload across asymmetric cores.In this paper,we exploit asymmetry of AMC architectures to allocate and schedule tasks.Based on MIT Cilk runtime system,SAWS consists of an asymmetric-aware task allocator and a selective work-stealing scheduler.The asymmetric-aware task allocator statically analyses parallel programs and adds runtime codes during the compiling phase.Then,during the execution of parallel programs and in the beginning of each iteration,task allocator manages to distributes the tasks properly to asymmetric CPUs so that most tasks can access data from local memory node and the workload is balanced according to the computational ability of different CPUs.After that,the selective work-stealing scheduler takes advantage of cores' power slack to further balance the workload at runtime or adjust the frequencies of asymmetric cores.We evaluate SAWS runtime system on a machine of AMC architecture,and our real-system experimental results show that SAWS improves the performance of memory-bound programs up to 59.3% compared with traditional work-stealing schedulers in AMC architectures without increasing the power consumption.In addition,in our scalability evaluation,SAWS also presents outstanding scalability.