Collaborative Research: Programming Models, Compilers, and Runtimes for High-End Computing on Manycore Processors

由于多核心技术的出现，高端计算(HEC)领域正处于一个重要的十字路口-在单个芯片上集成数十个处理器(核)，在2015-2020年的时间框架内，亿级系统中的每个芯片将达到1000个核。与前几代硬件演进不同，硬件路线图的这种转变将对未来的港灯软件产生深远影响。首先，程序员将面临在单个系统中表达和管理数百万到10亿线程规模的并行性的可伸缩性挑战。其次，程序员将面临在高度不统一的系统结构中优化数据移动的局部性挑战，在内存层次结构的相邻级别之间存在带宽和延迟的数量级差距。该项目的具体重点是开发编程模型、编译器和运行时来应对未来HEC系统的上述挑战，并以特定的多核体系结构为指导，以确保实用性。这项研究将在以下方面取得成果：1)新的多核体系结构的并行编程构造，如异步活动、位置和阶段，它们建立在过去使用X10语言的经验基础上，但在本研究中将用C/C++而不是JAVA来表现；2)新的并行中间表示(PIR？S)和编译器对并行性和局部性的优化；3)具有两级并行性的新的线程虚拟机，同步粗粒度线程(SCTS)和异步细粒度线程(AFTS)；4)在CyclopsC64多核系统上对编程模型、编译器和运行时进行了评估，这直接体现了未来HEC系统面临的并行性和局部性挑战。

The High End Computing (HEC) field is at a major crossroads due to the advent of many-core technology --- integration of tens and hundred processors (cores) on a single chip, heading to 1000 cores per chip in Exascale systems in the 2015-2020 timeframe. Unlike previous generations of hardware evolution, this shift in the hardware road-map will have a profound impact on future HEC software. First, the programmer will be faced with the scalability challenge of expressing and managing parallelism at the scale of millions to one billion threads in a single system. Second, the programmer will be faced with the locality challenge of optimizing data movement in a highly non-uniform system structure with order-of-magnitude gaps in bandwidth and latency between adjacent levels of the memory hierarchy.The specific focus of this project is on developing programming models, compilers, and runtimes to address the above challenges of future HEC systems, guided by a specific many-core architecture to ensure practicality. The research will deliver results in the following areas: 1) new parallel programming constructs for many-core architectures such as asynchronous activities, places, and phasers that build on past experiences with the X10 language, but will be manifested in C/C++ instead of Java in this research; 2) new parallel intermediate representations (PIR?s) and compiler optimizations for parallelism and locality; 3) a new thread virtual machine with two levels of parallelism, Synchronous Coarse-Grain Threads (SCTs) and Asynchronous Fine-Grain Threads (AFTs); and 4) evaluation of the programming model, compiler, and runtime on a Cyclops C64 many-core system that directly exemplifies the parallelism and locality challenges facing future HEC systems.