New sources of “big data” such as the Internet, mobile applications, data-driven science and large-scale sensors (IoT) are driving the demand for growing computing performance. The efficient analysis of data in native raw formats in real-time is increasingly important due to rapid data generation, popular demands for analytics and quick insights to derive immediate responses. Traditional data processing systems can deliver high performance on loaded data, but transforming raw data into internal system formats is computationally expensive. Data transformations rather than arithmetic operations dominate the task, and such transformation is a critical barrier to increased data processing performance. We propose a two-part software/hardware approach, namely, the unified transformation accelerator (hardware) and accelerated transformation operators (software), to accelerate data analytics on unloaded raw data and exceed the performance of structured data analytical systems. Together, they enable real-time decision making and fast knowledge exploration on messy, heterogeneous, and ad-hoc raw data. The unified transformation accelerator (UTA) creates new and flexible architecture support for analytical workloads. Exploiting customization and local memory for efficiency and MIMD parallelism for flexibility and performance, we designed a novel hardware architecture, the Unstructured Data Processor (UDP). UDP is a programmable general-purpose data transformation accelerator, customized for data analytics. We completed the instruction set architecture design, the micro-architecture implementation, the software toolchain (compiler, linker, loader), and the simulation infrastructures. The UDP has four unique features for performance and generality: multi-way dispatch, variable-size symbol, flexible-source dispatch, and flexible addressing. Extensive evaluation of data transformation kernels, which ranges from compression to pattern matching, shows UDP achieves 20x average speedup and 1900x energy efficiency when compared with an 8-thread CPU processor. The UDP ASIC implementation is > 100x less power and area than a single CPU core. The accelerated transformation operators (ATO) is a software framework for integrating hardware accelerations into data analytical systems in a pervasive and flexible fashion. It allows arbitrary data representations on query plan edges and arbitrary use of hardware acceleration within any operator implementations. This approach preserves full operator composition and query optimization capabilities, and requires essentially no front end and execution engine changes. We build the ACCORDA (Accelerated Operators for Raw Data Analysis) system, by extending the state-of-art distributed analytical system (SparkSQL) with the ATO approach. We will study the performance benefit of ACCORDA over a traditional analytical system (SparkSQL) using real-world end-to-end query workloads. We expect the unloaded raw data processing speed of the ACCORDA system to match or even exceed SparkSQL with loaded data. Furthermore, we describe the research plan and the full thesis outline in the proposal. The complete experiment results will be presented in the final thesis.
更多
分享
引用
