III: Small: Enabling the Best Utilization of GPUs for In-Memory Data Management Systems

对快速实时数据分析的需求不断增长，这就要求内存数据库系统同时提供高性能事务处理和低延迟查询响应时间。由于不同的工作负载特征和不同的服务质量要求，多个工作负载在传统的CPU/多核上共同运行，因此实现这两个目标是一项极具挑战性的任务。通用图形处理单元（GPU）的兴起为最大化查询执行性能同时最小化对事务处理的干扰提供了解耦机会。然而，为了在数据库系统中充分利用这种硬件设备，必须有效地解决面向并行计算的GPU架构与面向数据处理的数据库查询工作负载之间的一些不匹配问题。该项目寻求构建基于gpu的查询执行引擎的解决方案，并提供内存数据库系统作为前所未有的机会和能力，以高性能和低成本的方式执行事务工作负载和分析工作负载。本项目试图通过将基础研究成果转化为实用的数据库系统，通过研究活动训练本科生和研究生，并及时将新的研究成果引入课堂，从而产生更广泛的影响。具体而言，该项目将解决三个不匹配问题，包括：(1)GPU的高并行计算能力与从主存到主存的数据传输速度缓慢之间的不匹配；(2) GPU有限的编程能力与数据库查询复杂的结构之间的矛盾；(3)大量并行处理核心的可用性与缺乏对并发任务执行和内存管理的系统软件支持之间的关系。本项目将采用整体系统设计方法，通过开展几个密切相关的研究任务来解决这些挑战。(1)项目团队将开发一个具有多层次抽象和优化的软件翻译框架，将结构化查询语言查询自动翻译成在gpu上运行的高效程序。(2)考虑GPU性能因素，开发查询执行成本模型，支持查询优化和动态再优化。(3)团队将设计特定的算法和查询执行技术，以有效地处理两种复杂但实际重要的查询，即嵌套子查询和递归查询。(4)团队将构建用于系统级设备内存管理和动态查询调度的软件层。(5)最后的任务将是将项目成果集成到具有全面工作负载的代表性开源内存系统中。这个项目的研究成果将是开源的。

The continuously increasing demand for fast and real-time data analytics requires that in-memory database systems provide both high-performance transaction processing and low-latency query response time. To accomplish these two goals is a highly challenging task as multiple workloads are co-run on conventional CPU/multicores due to different workload characteristics and different quality of service requirements. The rise of general-purpose graphical processing units (GPU) brings a decoupling opportunity of maximizing query execution performance while minimizing interference to transaction processing. However, best utilization of such a hardware device in database systems must effectively address several mismatches between the parallel computing-oriented architecture of GPU and the data processing-oriented database query workloads. This project seeks a solution of building a GPU-based query execution engine and offers in-memory database systems as an unprecedented opportunity and capability to execute both transaction workloads and analytics workloads in a high-performance and low-cost way. This project attempts to make a high broader impact by transforming basic research results into practical database systems, and by training both undergraduate and graduate students with research activities, and by timely introducing new research results into classrooms.Specifically, the project will address three mismatching issues, including (1) the one between high parallel computing power in GPU and slow data transfer speeds from/to main memory; (2) the one between GPU's limited programming capabilities and database query's complex structures; and (3) the one between the availability of massive parallel processing cores and the lack of system software support for concurrent task executions and memory management. This project will apply a holistic system design methodology to cohesively address these challenges by carrying out several closely related research tasks. (1) The project team will develop a software translation framework with multi-level abstractions and optimizations to automatically translate Structured Query Language queries into highly-efficient programs running on GPUs. (2) The team will develop query execution cost models by considering GPU performance factors to support query optimization and dynamic re-optimization. (3) The team will design specific algorithms and query execution techniques to efficiently handle two complex but practically important queries, namely nested sub-queries and recursive queries. (4) The team will build a software layer for system-level device memory management and dynamic query scheduling. (5) A final task will be to integrate the project outcome into representative open source in-memory systems with comprehensive workloads. The research efforts in this project will be open source.