23 September 2020 in The VLDB Journal
I/O latency and throughput are two of the major performance bottlenecks for disk-based database systems. Persistent memory (PMem) technologies, like Intel’s Optane DC persistent memory modules, promise to bridge the gap between NAND-based flash (SSD) and DRAM, and thus eliminate the I/O bottleneck. In this paper, we provide the first comprehensive performance evaluation of PMem on real hardware in terms of bandwidth and latency. Based on the results, we develop guidelines for efficient PMem usage and four optimized low-level building blocks for PMem applications: log writing, block flushing, in-place updates, and coroutines for write latency hiding.
1 August 2020 in Proceedings of the VLDB Endowment
Relational database systems are purpose-built for a specific storage device class (e.g., HDD, SSD, or DRAM). They do not cope well with the multitude of storage devices that are competitive at their price `sweet spots'. To make use of different storage device classes, users have to resort to workarounds, such as storing data in different tablespaces. A lot of research has been done on heterogeneous storage frameworks for distributed big data query engines. These engines scale well for big data sets but are often CPU- or network-bound. Both approaches only maximize performance for previously purchased storage devices. We present Mosaic, a storage engine for scan-heavy workloads on RDBMS that manages devices in a tierless pool and provides device purchase recommendations for a specified workload and budget. In contrast to existing systems, Mosaic generates a performance/budget curve that is Pareto-optimal, along which the user can choose. Our approach uses device models and linear optimization to find a data placement solution that maximizes I/O throughput for the workload. Our evaluation shows that Mosaic provides a higher throughput at the same budget or a similar throughput at a lower budget than the state-of-the-art approaches of big data query engines and RDBMS.