DEMETER: hardware-assisted database checkpointing

SAC '20: The 35th ACM/SIGAPP Symposium on Applied Computing Brno Czech Republic March, 2020(2020)

引用 0|浏览23
暂无评分
摘要
For the past four decades, database checkpointing has long been a major system activity triggering heavy I/O and leading to performance fluctuations. This is unpleasing to database systems pursuing good and steady performance on high-end computing platforms. Hence, commercial and open-source databases usually spawn massive page flushers to complete the required tasks with system disruptions being minimized. In this work, to reduce the hefty cost, we propose Demeter, designs for hardware-assisted database checkpointing that unbundles checkpointing by offloading the task of building new pages from physical changes onto storage devices equipped with near-data processing (NDP) capability. Demeter consists of two parts; a software component designed for generating physical updates without affecting foreground transactions and a hardware part constructing new pages from the changes. For evaluation, we have built a prototype system based on the Biscuit framework and integrated Demeter with MariaDB. Experimental results demonstrated that the Demeter-enabled engine substantially reduces the amount of I/O and its foreground transaction processing gains more throughput than the vanilla system under the OLTP workloads modifying a large volume of data pages.
更多
查看译文
关键词
database checkpoint, near-data processing
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要