Author : Tianzheng Wang
Publisher :
ISBN 13 :
Total Pages : pages
Book Rating : 4.:/5 (133 download)
Book Synopsis Fast and Robust Transaction Processing on Emerging Hardware by : Tianzheng Wang
Download or read book Fast and Robust Transaction Processing on Emerging Hardware written by Tianzheng Wang and published by . This book was released on 2018 with total page pages. Available in PDF, EPUB and Kindle. Book excerpt: Database engines must adapt to the underlying hardware for high-performance transaction execution. Conventional database engines are built on the storage-centric architecture assuming small DRAM, limited parallelism, and slow network. Emerging hardware, including byte-addressable non-volatile memory (NVRAM), massively parallel processors, and ultra-fast interconnect, has lead to significant changes in the design and implementation of modern database engines. Modern database engines, despite being optimized for multicore hardware, still have unscalable components and unnecessary data redundancy that hinder performance on future servers with large core counts (up to 1000). Database engines should also be robust against various workloads, including traditional short, write-intensive transactions and emerging long, read-mostly, and high-conflict transactions. Existing database engines fail to address these problems, leaving the vast opportunity of accommodating heterogeneous, read-mostly and high-conflict workloads that are enabled by emerging hardware largely unexplored. These problems must be tackled to make database engines perform fast and robustly on modern high-end and future hardware. In this thesis, we first remove the unscalable logging bottleneck by utilizing NVRAM. NVRAM allows us to revive distributed logging, a once prohibitive but effective solution to centralized logging bottlenecks. Avoiding unscalable logging is crucial in building fast main-memory databases that can potentially scale up to 1000 cores in future servers, but is not enough for handling the inconvenient workloads robustly. We continue on devising concurrency control mechanisms that handle a wide spectrum of workloads robustly. Finally, we demonstrate how high availability solutions can be built to guarantee both data safety and freshness using emerging hardware and by eliminating unnecessary data redundancy through indirection. This thesis serves as a toolbox and provides the guidelines for the design and implementation of high-performance and robust database engines and other concurrent systems. The solutions have been evaluated on both mainstream high-end servers and precursors of future hardware that could feature as many as 1000 cores and NVRAM. The techniques are applicable to both modern and future hardware that is yet to come.
