CAREER: Flexible Architectures for Data-Intensive Computing

这项研究的重点是数据密集型计算-在传统计算机系统的内存和I/O子系统中嵌入处理能力，并将代码传送到数据中，而不是通过主微处理器移动数据。可以利用存储器和I/O系统中的这种新发现的灵活性的新型体系结构具有显著提高性能和健壮性的潜力。它们还将消除当前并行和单处理器体系结构中的三个重要限制：可扩展多处理器中的硬连线高速缓存一致性协议、跨距访问较差或空间局部性较差的应用程序的较差高速缓存性能，以及I/O密集型应用程序中的延迟和带宽瓶颈。这项研究专门解决了这些限制，包括为可编程存储系统开发新的一致性协议和公平机制；以及开发必要的协议扩展以将闪存多处理器转换为活动存储器机。这将通过利用硬件原型的存在来运行真实世界的问题并深入了解灵活内存和I/O系统的潜力来实现。最后，完成了面向数据密集型计算的高性能新型统一架构Active Fabric的设计。

This research focuses on data-intensive computing---embedding processing capability in the memory and I/O subsystems of traditional computer systems and shipping the code to the data rather than moving the data through the main microprocessor. Novel architectures that can exploit this newfound flexibility in the memory and I/O system have the potential to significantly improve performance and robustness. They will also remove three important limitations in current parallel and uniprocessor architectures: hardwired cache coherence protocols in scalable multiprocessors, poor cache performance for applications with bad stride accesses or poor spatial locality, and latency and bandwidth bottlenecks in I/O-intensive applications. This research specifically addresses those limitations and includes the development of novel coherence protocols and fairness mechanisms for programmable memory systems; and the development of the necessary protocol extensions to convert the FLASH multiprocessor into an active memory machine. This will be done by exploiting the existence of a hardware prototype to run real-world problems and gain insight into the potential of flexible memory and I/O systems. Finally, the design of Active Fabric, a high performance novel unified architecture for data-intensive computing will be accomplished.