Performance Overview

In version 9.1.0, we implemented many performance optimization features to improve the overall SQL query performance. We tested the performance of the latest version 9.1.0 against version 8.3.0 using TPC-H and TPC-DS benchmarks with a 1 TB dataset. The cluster had 6 nodes, each with 16 vCPUs and 64 GB specifications, totaling 96 vCPUs and 384 GB. The results showed that:

• The TPC-H total query time for the 9.1.0 version with the storage-compute coupled architecture was 170.08 seconds, a 208% improvement over the 533.05 seconds of the 8.3.0 version.
• In 9.1.0, the performance degradation of the storage-compute decoupled architecture and the coupled architecture was within 10%.
• In the TPC-H 1000x benchmark, the query performance of all 22 SQL statements in the 9.1.0.200 version showed performance improvements ranging from 1.5 to 5 times compared to the 8.3.0 version. Specifically, Q19 showed a remarkable 16-fold improvement.
• Similarly, in the TPC-DS 1000x benchmark, the 9.1.0.200 version demonstrated significant improvements in 87 out of 99 SQL queries compared to the 8.3.0 version, with 15 queries showing performance gains of 2 to 10 times.
• The 9.1.0 version also showcased notable advantages in various operations such as filtering, sorting, aggregation, multi-table joins, window calculations, and CTE queries.

In version 9.1.0, a storage-compute decoupling architecture is utilized to specify level-2 partitions. The performance comparison between GaussDB(DWS) and ClickHouse is based on the ssb-flat 100 GB test benchmark. The results showed that:

• The initial performance is 200% higher than that of the open source vendor ClickHouse.
• The initial performance is improved by 133% compared with version 9.1.0.100.