基于最佳并行度的任务依赖图调度

基于最佳并行度的任务依赖图调度策略充分利用编译时刻所得到的全局信息，采用横向和纵向任务合并，处理节点预分配，静态调度和动态调度相结合、集中式调度和分层调度相结合等措施，是一种简单的、具有较高效率的实用化调度方案．该调度方案能够在尽量压缩调度长度的情况下节约系统资源．

DYNAMEM-A Microarchitecture for Improving Memory Disambiguation at Run-Time

This paper presents a new microarchitecture technique named DYNAMEM,in which memory reference instructions are dynamically scheduled and can be executed out-of-order. Load instructions can bypass store instructions specula-tively, even if the store instructions'addresses are unknown. DYNAMEM can greatly alleviate the restraints of ambiguous memory dependencies. Simulation results show that the frequency of false load is low. Mechanism has been pro-vided to repair false loads with low penalty, and to achieve precise interrupts.Discussions and experimental results show that DYNAMEM could dramatically raise instruction-level parallelism in programs without recompilation.

基于共享内存的并行LDA算法

现有的共享内存的并行潜在狄利克雷分配(LDA)主题模型,通常由于数据分布的原因,线程之间一般存在等待导致效率低下。针对线程等待问题进行研究,提出一种基于动态的线程调度方案。该方案能够根据线程的数量进行分块,在此基础上及时为空闲的线程动态地分配任务,从而减少线程间等待时间。实验表明,这种新的调度方案能够有效地解决线程等待问题。该方案不仅在保证收敛精度的同时能够获得加速比25%的提升,还能显著提高向上扩展比。对于大规模分布式集群上单个节点的并行LDA算法来说,这种调度可以更有效地利用计算资源。