Scan and Join Optimization by Exploiting Internal Parallelism of Flash-Based Solid State Drives

Abstract

Abstract. Nowadays, flash-based solid state drives (SSDs) are gradually replacing hard disk drives (HDDs) as the primary non-volatile storage in both desktop and enterprise applications because of their potential to speed up performance and reduce power consumption. However, database query processing engines are designed based on the fundamental characteristics of HDDs, so they may not benefit immediately from SSDs. Previous researches on optimizing database query processing on SSDs have mainly focused on leveraging the high random data access performance of SSDs and avoiding slow random writes whenever possible. However, they fail to exploit the rich internal parallelism of SSDs. In this paper, we focus on exploiting rich internal parallelism of SSDs to optimize scan and join operators. Firstly, we detect internal parallelism of SSDs seemed as black boxes. Then we propose a parallel table scan operator called ParaScan to take full advantage of internal parallelism of SSDs. Based on ParaScan, we also present an efficient parallel join operator called ParaHashJoin to accelerate database query processing. Experimental results on TPC-H datasets show that our ParaScan on SSD significantly outperforms the traditional table scan on SSD by 1X, and ParaHashJoin is 1.5X faster than traditional hash join operator especially when join selectivity is small.

Exploiting Internal Parallelism of SSD for Hash Join

“…Experimental results demonstrate that the idea of CEC can be used to adjust the behavior of the hash join algorithm in response to the changing status of execution environment, and that the proposed buffer allocation scheme works well and can ensure high throughput without wasting memory resources. By the way, please note that our black box approach can naturally exploits the advantage of those white box approaches [9,19,13,12,20] . For example, if data is placed in two different channels, its data access parallelism can be used by our approach.…”

“…Roh et al [20] proposed a scheme called psync I/O, which enables the usage of SSDs' internal parallelism within a single process, and propose a new variant of B+-tree called PIO Btree by incorporating psync I/O into a B+-tree. Lai et al [13] optimize the scan and join operations by exploiting SSDs' internal parallelism, where they firstly detect SSDs' internal structure and data placement strategy by using the methods described in Ref. [7] and place the data accordingly, and then use the internal parallelism to access data by directly addressing the physical address when executing scan or join operations.…”

“…For each TIF, we calculate the corresponding parameters for GHJ as indicated in Section IV.2. For example, for CUSTOMER table, given a TIF value 2 13 , if the degree of parallelism T is set to 16, then the block size will be β = 2 13 /16=512KB, |IB| = 4β=2MB, |BB| = |P B| = |OB| = 2β=1MB, and the required memory size M = √ 2β × F × T × |R| + 4βT is around 172MB. In our experiment, we define CEC by setting the standard deviation threshold to 10.…”

“…According to Fig.8, the experimental results are consistent with the SSDs' throughput results when applying CEC. Take SSD-850 as an example, when TIF is larger than 2 13 , the average performance (time) of GHJ almost reaches the optimal performance and the standard deviation is less than 10. This indicates that, within each CEC (with a specific TIF value), the combinations of queue depth and I/O block size can substitute each other without any significant difference in performance.…”

“…This means that parallelism exists at various levels, including plane-, die-, flash package-and channel-levels. To boost the performance of SSD-based DBMSs, some researchers have worked on obtaining SSDs' internal skeleton structure and rules of data organization via designing special test cases and then exploited them to improve the access performance of SSDbased DBMSs [9,13] .…”