Flash-aware index scan in PostgreSQL

Recently, the trend of storage markets has changed from HDDs that have dominated the markets for the last several decades to flash based SSDs. Corresponding to the drift, various studies have been conducted to adapt traditional DBMS to SSD based storage devices. However, most DBMSs are still more HDD-friendly. HDDs and SSDs have inherent features because of their own architecture designs. HDD has a wide gap between the performance of sequential I/O and that of random I/O owing to its mechanical parts. Due to this fact, DBMS usually prefer a full table scan to an index scan except when a selectivity is enough low to take advantage of index scan. Unlike HDDs, SSDs do not contain any mechanical parts and therefore there is a tiny gap between the performance of sequential I/O and that of random I/O. This characteristic of SSDs allows DBMSs to efficiently access a storage based SSD with an index scan. Another feature of SSDs is an internal parallelism thanks to its internal architecture. In spite of a circumstance with the secondary storage based SSD, DBMS are more likely to choose a full table scan rather than an index scan for I/O operations. It is necessary to understand the distinct properties and differences between the two storage devices and make index scans SSD-friendly to improve their performance. In this paper, we implement an optimization of the index scan, called flash-aware index scan in the paper, by combining two concepts in PostgreSQL. A concept is a sorted index scan that scans tuples in order of record ids. The other is parallel synchronous I/O that is a traditional synchronous I/O with an array of I/O requests per operation. The goal of the paper is making the system aware of the nature of SSDs, thus enhancing the performance of the index scans.