Highlighting the performance diversity of analytical queries using VOILA

Hardware architecture has long influenced software architecture, and notably so in analytical database systems. Currently, we see a new trend emerging: A "tectonic shift" away from X86-based platforms. Little is (yet) known on how this shift affects database system performance and, consequently, should influence the design choices made. In this paper, we investigate the performance characteristics of X86, POWER, ARM and RISC-V hardware on micro- as well as macro-benchmarks on a variety of analytical database engine designs. Our tool to do so is VOILA: a new database engine generator framework that from a single specification can generate hundreds of different database architecture engines (called "flavors"), among which well-known design points such as vectorized and data-centric execution. We found that performance on different queries by different flavors varies significantly, with no single best flavor overall, and per query different flavors winning, depending on the hardware. We think this "performance diversity" motivates a redesign of existing – inflexible – engines towards hardware- and query-adaptive ones. Additionally, we found that modern ARM platforms can beat X86 in terms of overall performance by up to 2×, provide up to 11.6× lower cost per instance, and up to 4.4× lower cost per query run. This is an early indication that the best days of X86 are over