Alternative Titleビットマップインデックスに基づくデータ解析のためのハードウェアシステムに関する研究
Note (General)Recent years have witnessed a massive growth of global data generated from web services, social media networks, and science experiments, as well as the “tsunami" of Internet-of-Things devices. According to a Cisco forecast, total data center traffic is projected to hit 15.3 zettabytes (ZB) by the end of 2020. Gaining insight into a vast amount of data is highly important because valuable data are the driving force for business decisions and processes, as well as scientists' exploration and discovery.To facilitate analytics, data are usually indexed in advance. Depending on the workloads, such as online transaction processing (OLTP) workloads and online analytics processing (OLAP) workloads, several indexing frameworks have been proposed. Specifically, B+-tree and hash are two common indexing methods in OLTP, where the number of querying and updating processes are nearly similar. Unlike OLTP, OLAP concentrates on querying in a huge historical storage, where updating processes are irregular. Most queries in OLAP are also highly complex and involve aggregations, while the execution time is often limited. To address these challenges, a bitmap index (BI) was proposed and has been proven as a promising candidate for OLAP-like workloads.A BI is a bit-level matrix, whose number of rows and columns are the length and cardinality of the datasets, respectively. With a BI, answering multi-dimensional queries becomes a series of bitwise operators, e.g. AND, OR, XOR, and NOT, on bit columns. As a result, a BI has proven profitable for solving complex queries in large enterprise databases and scientific databases. More significantly, because of the usage of low-hardware logical operators, a BI appears to be suitable for advanced parallel-processing platforms, such as multi-core CPUs, graphics processing units (GPUs), field-programmable logic arrays (FPGAs), and application-specific integrated circuits (ASIC).Modern FPGAs and ASICs have become increasingly important in data analytics because they can confront both data-intensive and computing-intensive tasks effectively. Furthermore, FPGAs and ASICs can provide higher energy efficiency, compared to CPUs and GPUs. As a result, since 2010, Microsoft has been working on the so-called Catapult project, where FPGAs were integrated into datacenter servers to accelerate their search engine as well as AI applications. In 2016, Oracle for the first time introduced SPARC S7 and M7 processors that are used for accelerating the OLTP databases. Nonetheless, a study on the feasibility of BI-based analytics systems using FPGAs and ASICs has not yet been developed.This dissertation, therefore, focuses on implementing the data analytics systems, in both FPGAs and ASICs, using BI. The advantages of the proposed systems include scalability, low data input/output cost, high processing throughput, and high energy efficiency. Three main modules are proposed: (1) a BI creator that indexes the given records by a list of keys and outputs the BI vectors to the external memory; (2) a BI-based query processor that employs the given BI vectors to answer users' queries and outputs the results to the external memory; and (3) an BI encoder that returns the positions of one-bits of bitmap results to the external memory. Six hardware systems based on those three modules are implemented in an FPGA in advance for functional verification and then partially in two ASICs|180-nm bulk complementary metal-oxide-semiconductor (CMOS) and 65-nm Silicon-On-Thin-Buried-Oxide (SOTB) CMOS technology―for physical design verification. Based on the experimental results, these proposed systems outperform other CPU-based and GPU-based designs, especially in terms of energy efficiency.
2017
Collection (particular)国立国会図書館デジタルコレクション > デジタル化資料 > 博士論文
Date Accepted (W3CDTF)2017-12-04T02:02:48+09:00
Data Provider (Database)国立国会図書館 : 国立国会図書館デジタルコレクション