An Experimental Study across GPU DBMSes toward Cost-Effective Analytical Processing

Abstract

Young-Kyoon SUH †a) , Member, Seounghyeon KIM †b) , Joo-Young LEE †c) , Hawon CHU †d) , Junyoung AN †e) , and Kyong-Ha LEE † †f) , Nonmembers SUMMARYIn this letter we analyze the economic worth of GPU on analytical processing of GPU-accelerated database management systems (DBMSes). To this end, we conducted rigorous experiments with TPC-H across three popular GPU DBMSes. Consequently, we show that co-processing with CPU and GPU in the GPU DBMSes was cost-effective despite exposed concerns.

Workload-Driven Analysis on the Performance Characteristics of GPU-Accelerated DBMSes

“…Copyright c 2022 The Institute of Electronics, Information and Communication Engineers v5.5.0. As mentioned in our previous work [1], these systems were chosen due to their distinguished strengths: e.g., executable code provision, open-sourced code, and relational query support. We used the TPC-H benchmark [13] as our dataset.…”

“…On the same database, BlazingSQL, OmniSciDB, and PG-Strom outperformed PostgreSQL by 94x, 32x, and 3x, respectively, for Q12-4, demanding a four-way join followed by sorting (caused by order-by). However, the studied GPU DBMSes are not recommended for join and simple selective scan queries on small data that easily fit in GPU mem-ory, and they may not be cost-effective as discussed in our earlier research [1].…”

“…Recently, GPU database management systems (DBMSes) have gained popularity due to the query processing acceleration on GPUs by leveraging a high degree of parallelism [1]- [3]. Many business OLAP applications can benefit from the GPU-accelerated execution of complex analytical tasks in DBMSes.…”