XPS: Full: CCA: Enhancing Scalability and Energy Efficiency in Extreme-Scale Parallel Systems through Application-Aware Communication Reduction

目前可用的高性能计算系统远远不能满足未来计算系统所需的性能和能源效率目标。随着处理器时钟频率在过去几年中趋于平稳，在相同的物理尺寸和相同或更低的能量需求下实现预期性能水平的唯一方法是让更多的处理器并行计算。然而，增加处理器之间的并发性会增加处理器间的通信，从而成为关键的瓶颈。这项工作提出了一种以应用程序为中心的方法，以最大限度地减少通信引起的开销，从而实现极端的可伸缩性。可以利用未来并行应用程序的隐式容错性来减轻高度并发系统的局域性退化和增加的通信成本，同时确保可接受的应用程序输出质量水平。这项工作特别关注(i)如何压缩和可能丢弃传输的数据，以及（ii）如何在某些应用程序上下文中避免或放松通信和同步，同时保持特定应用程序可接受的输出质量水平。这项工作的智力价值源于它对跨层通信和同步技术的探索，该技术利用应用程序的特性来显着降低通信和同步的成本。这种方法将使更广泛的应用程序能够利用并行性，从而提高它们的可伸缩性。这项工作将通过量化与应用感知通信和同步相关的成本和收益，设计应用感知可扩展通信和同步原元，开发支持应用感知可扩展通信的软件，以及为应用感知可扩展通信设计新的微架构和系统支持，从而实现在大规模计算系统上执行的特定领域应用的极端可扩展性。这项研究更广泛的影响在于显著降低了广泛应用程序的通信和同步成本，从而使一系列应用程序的并行可扩展性和能源效率达到新的水平。该项目将为分析和增强并行应用程序的可扩展性提供开源工具，并将为研究生和本科生提供参与尖端计算机系统研究的新机会。

The high-performance computing systems available today are far from meeting the performance and energy efficiency targets necessary to satisfy the needs of future computing systems. With processor clock frequencies leveling off over the past several years, the only way to achieve the expected levels of performance within the same physical size and the same or lower energy requirements is to have more processors computing in parallel. However, increasing concurrency among the processors increases inter-processor communication, which then becomes the critical bottleneck. This work proposes an application-centric approach to minimize communication-induced overheads in order to enable extreme scalability. The implicit fault tolerance of future parallel applications can be exploited to mitigate the degraded locality and increased communication costs of highly concurrent systems while ensuring an acceptable level of application output quality. This work specifically focuses on (i) how data to be transferred can be compressed, and possibly discarded, and (ii) how communication and synchronization can be avoided or relaxed in certain application contexts, all while maintaining an acceptable level of output quality for the specific application. The intellectual merit of this work stems from its exploration of cross-layer communication and synchronization techniques that exploit application characteristics to significantly reduce the cost of communication and synchronization. This approach will enable a wider array of applications to exploit parallelism and thereby increase their scalability. This work will enable extreme scalability of domain-specific applications executing on large-scale computing systems by quantifying the costs and benefits associated with application-aware communication and synchronization, designing application-aware scalable communication and synchronization primitives, developing software to support application-aware scalable communication, and designing new micro-architectural and system supports for application-aware scalable communication.The broader impact of this research lies in significantly reducing the costs of communication and synchronization for a wide spectrum of applications and consequently enabling new levels of parallel scalability and energy efficiency across a range of application programs. This project will produce open source tools for analyzing and enhancing the scalability of parallel applications, and will provide new opportunities for both graduate and undergraduate students to participate in cutting-edge computer systems research.