CAREER: In-Situ Compute Memories for Accelerating Data Parallel Applications

由于当今的计算是由大数据主导的，因此对这一重要领域的专业化有着强烈的推动作用。这些以数据为中心的应用程序的性能在很大程度上取决于数据的高效访问和处理。这些应用程序往往是高度数据并行的，需要处理大量数据。最近的研究表明，到2020年，来自个人和公司的数据产量预计将增长到73.5ZB，比2015年增长4.4倍。此外，与实际计算相比，它们倾向于在将数据从存储移动到计算单元以及在指令处理上花费不成比例的大比例的时间和精力。这项研究寻求设计专门的以数据为中心的计算系统，以显著减少这些开销。在通用计算系统中，大部分聚集裸片区域(超过90%)专门用于在存储器层次结构中的几个级别存储和检索信息：片上高速缓存、主存储器(DRAM)和非易失性存储器(NVM)。这项研究的中心愿景是创造原位计算存储器，它重新调整这些存储结构中使用的元素的用途，并将它们转换为活动的计算单元。与在存储器方法中增加存储器阵列外部的逻辑的先前处理不同，原位计算存储器背后的基础原理是允许在每个存储器阵列内就地计算，而不将数据传送进或传送出它。这样的转换可以释放海量数据并行计算能力(高达100倍)，并减少通过各级内存层次结构(高达20倍)移动数据所花费的能量，从而直接满足以数据为中心的应用程序的需求。这项工作开发了现场计算内存技术，调整了系统软件堆栈，并重新设计了以数据为中心的应用程序，以利用这些内存。

As computing today is dominated by Big Data, there is a strong impetus for specialization for this important domain. Performance of these data-centric applications depends critically on efficient access and processing of data. These applications tend to be highly data-parallel and deal with large amounts. Recent studies show that by the year 2020, data production from individuals and corporations is expected to grow to 73.5 zetabytes, a 4.4× increase from the year 2015. In addition, they tend to expend disproportionately large fraction of time and energy in moving data from storage to compute units, and in instruction processing, when compared to the actual computation. This research seeks to design specialized data-centric computing systems that dramatically reduce these overheads. In a general-purpose computing system, the majority of the aggregate die area (over 90%) is devoted for storing and retrieving information at several levels in the memory hierarchy: on-chip caches, main memory (DRAM), and non-volatile memory (NVM). The central vision of this research is to create in-situ compute memories, which re-purpose the elements used in these storage structures and transform them into active computational units. In contrast to prior processing in memory approaches, which augment logic outside the memory arrays, the underpinning principle behind in-situ compute memories is to enable computation in-place within each memory array, without transferring the data in or out of it. Such a transformation could unlock massive data-parallel compute capabilities (up to 100×), and reduce energy spent in data movement through various levels of memory hierarchy (up to 20×), thereby directly address the needs of data-centric applications. This work develops in-situ compute memory technology, adapts the system software stack and re-designs data-centric applications to take advantage of those memories.