DWS: Demand-aware Work-Stealing in Multi-programmed Multi-core Architectures

ABSTRACTTraditional work-stealing schedulers perform poorly in multi-programmed multi-core architectures, because all the programs tend to use all the cores and thus incur serious core contention. To relieve this problem, this paper proposes a Demand-aware Work-Stealing (DWS) task scheduler, with which a work-stealing program uses cores according to its realtime demand on the cores. If multiple programs scheduled by DWS run in a multi-core architecture concurrently, the cores are first evenly allocated to the co-running programs. At runtime, if a program cannot fully utilize its cores, it releases some of its allocated cores. Otherwise, if a program demands more cores, it tries to use the free cores released by its co-running programs. Experimental results show that DWS can achieve up to 32.3% performance gain for co-running programs compared to traditional work-stealing schedulers with the ABP yielding mechanism.