High-Performance Synchronization for Shared-Memory Parallel Programs

小行星9319445 本研究的目的是发展和评估超大规模共享内存多处理器系统的同步技术。 随着并行处理器的大小和可用性的增加，高性能同步已成为整体应用程序性能和可伸缩性的关键因素。 同步技术包括专用硬件、活动消息、无锁数据结构和基于原子取O指令的可扩展软件同步算法。 同步方法的选择对整体性能起着重要作用。 PI建议将他的研究集中在1）软件同步，无锁数据结构，活动消息和专用硬件原语的比较评估，2）随着处理器数量的增加而扩展并利用可用原子指令的新同步算法，以及3）用于合作同步和调度的新机制，其最大限度地减少了锁上的不必要旋转，最大化处理器局部性并避免对锁定和非锁定数据的争用。 ***

9319445 Scott The goal of this research is to develop and evaluate synchronization techniques on very large scale shared memory mutliprocessor systems. With increases in size and availability of parallel processors, high-performance synchronization has become a critical factor in overall application performance and scalability. Among the techniques for synchronization include special-purpose hardware, active messages, lock-free data structures, and scalable software synchronization algorithms based on atomic fetch-and-o instructions. The choice of the synchronization approach plays a significant role on the overall performance. The PI proposes to concentrate his research on 1)comparative evaluation of software synchronization, lock- free data structures, active messages and special-purpose hardware primitives, 2)new synchronization algorithms that scale with the number of processors and utilize the available atomic instructions, and 3) new mechanisms for cooperative synchronization and scheduling which minimize the unnecessary spinning on locks, maximize processor locality and avoid contention for both lock and nonlock data. ***