arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 1702.03008cs.DC
收藏

并发动态平面凸包中的细粒度锁

Finer-grained Locking in Concurrent Dynamic Planar Convex Hulls

K. Alex Mills, James Smith

摘要:

平面点集的凸包是包含点集内每个点的最小凸多边形。动态凸包问题涉及到有效地保持一组点的凸包的增加和删除。这个问题的一个算法使用两个外部平衡二叉搜索树(BST)(M。H. Overmars,J. van Leeuwen 1981)。我们提出了这个问题的第一个并发解决方案,它使用一个单一的BST存储引用中间凸包解决方案在每个节点。我们实现并评估了两种基于锁的方法:a)细粒度锁定,其中树的每个节点都受到锁的保护,以及b)“细粒度锁定”,其中每个节点包含用于左链和右链的单独锁。在我们的吞吐量实验中,我们观察到细粒度锁定比细粒度锁定提高了8-60%,比粗粒度锁定和软件事务内存(STM)提高了38- 61倍。当应用于寻找静态点集的凸包时,我们的方法比使用同等数量的线程的并行分治实现的性能高2- 4倍。

英文摘要:

The convex hull of a planar point set is the smallest convex polygon containing each point in the set. The dynamic convex hull problem concerns efficiently maintaining the convex hull of a set of points subject to additions and removals. One algorithm for this problem uses two external balanced binary search trees (BSTs) (M. H. Overmars, J. van Leeuwen 1981). We present the first concurrent solution for this problem, which uses a single BST that stores references to intermediate convex hull solutions at each node. We implement and evaluate two lock-based approaches: a) fine-grained locking, where each node of the tree is protected by a lock, and b) "finer-grained locking", where each node contains a separate lock for each of the left and right chains. In our throughput experiments, we observe that finer-grained locking yields an 8-60% improvement over fine-grained locking, and a 38-61x improvement over coarsegrained locking and software transactional memory (STM). When applied to find the convex hull of static point sets, our approach outperforms a parallel divide-and-conquer implementation by 2-4x using an equivalent number of threads.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑