SHF: Small: Non-Uniformity--Centric Program Optimizations for Dynamic Computations on Chip Multiprocessors

这一研究项目的动机是处理器发展趋势与现代数据密集型动态应用需求之间的日益扩大的差距，从微分方程式解算器到数据挖掘工具再到粒子动力学模拟，在科学和人文领域发挥着重要作用。这些应用程序具有海量的数据访问以及复杂的数据访问或控制流模式。这些特性使它们成为现代处理器面临的巨大挑战，而现代处理器的发展与这些应用的需求恰恰相反：芯片的聚合计算能力正在迅速超过内存带宽；面向吞吐量的多核的兴起使系统吞吐量对不规则计算更加敏感。以非一致性为中心的程序优化新范式独特地将现代系统中的非一致性核心关系作为程序优化的一阶约束。PipeReg框架是一种在运行时重新组织数据访问和线程的新方法，以减少不规则计算对大规模并行处理器吞吐量的影响。此外，在多插槽多核系统中，利用片上存储中线程间的非均匀交互，提出了一种新的邻域感知程序转换方法。这两种技术结合在一起，将协同消除一些重要的障碍，使数据密集型动态应用程序能够充分利用未来计算系统的能力。该研究成果将为提高数据密集型动态应用在异构型并行系统时代的计算效率提供必要的支持。由于这些应用的关键作用，这项研究将有助于促进科学、商业、卫生和其他领域的持续进步。

This research project is motivated by a growing gap between trends in processor development and needs of modern data-intensive dynamic applications ranging from differential equation solvers to data mining tools to particle dynamics simulations, playing an essential role in science and humanity. These applications feature tremendous amounts of data accesses as well as complex patterns in data accesses or control flows. The properties make them a great challenge for modern processors which are evolving exactly opposite to these applications' needs: A chip's aggregate computing power is rapidly outgrowing memory bandwidth; the rise of throughput-oriented manycores makes system throughput even more sensitive to irregular computations. The new paradigm of program optimizations, non-uniformity--centric, distinctively takes the non-uniform inter-core relations in modern systems as the first-order constraint for program optimizations. The proposed framework named PipeReg is a new way to reorganize data accesses and threads during run time to reduce the influence of irregular computations on the throughput of massively parallel processors. In addition, a novel kind of program transformations, neighborhood-aware, exploiting non-uniform interactions among threads in on-chip storage in a multi-socket multicore system. Together, the two techniques will synergistically remove some important barriers for data-intensive dynamic applications to tap into the full power of future computing systems. The outcome from this research will provide essential support for enhancing the computing efficiency of data-intensive dynamic applications in the era of heterogeneous parallel systems. Because of the critical roles of these applications, this research will help foster sustained advancement in science, commerce, health, and other areas.