Siaw: an Adaptive Idleness-Aware Work-Stealing Strategy on Shared Memory Machines

Lei Guan,Yu-tong LU,Tao Gao
DOI: https://doi.org/10.1109/isai.2016.0060
2016-01-01
Abstract:Work-stealing plays a key role in promoting load balancing for such applications that need continuous dynamic load balancing. However, the performance of current work-stealing algorithms is usually very sensitive to the choice of steal granularities since these algorithms typically adopt a fixed steal granularity. On the other hand, these work-stealing algorithms usually obtain poor performance for large steal granularities due to load imbalance. In this paper we present an adaptive idleness-aware work-stealing strategy named SIAW, aiming at promoting load balancing on shared memory machines. Using the unbalanced tree search (UTS) benchmark, we detail the implementation of our algorithm and report the experimental results. Compared with shared memory implementation of UTS benchmark, our approach has great advantage in promoting load balancing for large task granularities. Most importantly, the experimental results illustrate that our approach makes it possible to get good performance over a wider range of task granularities. When using 32 threads, we also observe a factor of 26% performance improvement, for the benchmark studied in this paper.
What problem does this paper attempt to address?