No Silver Bullet: Boosting GaussDB Performance on the 30TB TPC-H Workload

A pipeline execution model, scalable inter-node shuffle, unified bus, and cross-node Bloom filters deliver a 40% improvement over the best published TPC-H result at 30 TB.

Tim Zeyl; Jason Lam; Shu Lin; Reza Pournaghi; Qi Cheng; Calvin Wong; Kaixiang Du; Yuliang He; Yang Sun; Weicheng Wang; Paul Lee; Chen Ruo; Yang Xinyi; Li Qunan; Wang Junjie; Hu Dongxing; Chong Chen; Per-Ake Larson· 2026· DOI 10.48550/arXiv.2608.28352

The core problem

GaussDB is Huawei's premier database system, engineered for large-scale deployments and the most demanding workloads. It is a distributed shared-nothing system capable of handling all types of workloads. This paper outlines a series of modifications to GaussDB aimed at improving its performance on large-scale and complex analytical workloads. After these changes, its performance on the TPC-H workload exceeded the best published result by 40% at 30 TB. The key enhancements to achieve this elite performance include adopting a pipeline execution model, a faster and more scalable inter-node data shuffle, exploiting a unified bus and unified remote memory access. The authors also expanded the support of cost-based Bloom filter placement and implemented several Bloom filter streaming strategies, enabling their use across nodes.

Innovation

After the described modifications, GaussDB's performance on the TPC-H workload exceeded the best published result by 40% at 30 TB. This improvement is attributed to the combination of the pipeline execution model, the faster and more scalable inter-node data shuffle, the unified bus and unified remote memory access, and the expanded Bloom filter capabilities. The paper reports that these enhancements collectively enabled elite performance at this large scale. No further breakdown of individual component contributions is provided in the abstract.

GaussDB is Huawei's premier database system, engineered for large-scale deployments and the most demanding workloads. It is a distributed shared-nothing system capable of handling all types of workloads. This paper outlines a series of modifications to GaussDB aimed at improving its performance on large-scale and complex analytical workloads. After these changes, its performance on the TPC-H workload exceeded the best published result by 40% at 30 TB. The key enhancements to achieve this elite performance include adopting a pipeline execution model, a faster and more scalable inter-node data shuffle, exploiting a unified bus and unified remote memory access. The authors also expanded the support of cost-based Bloom filter placement and implemented several Bloom filter streaming strategies, enabling their use across nodes.

The authors implemented a set of architectural and algorithmic modifications to GaussDB. The core methodological pillars are:

Why it matters

The title "No Silver Bullet" suggests that the performance gains were not the result of a single optimization but rather the cumulative effect of multiple coordinated changes across the execution engine, data exchange layer, hardware utilization, and query optimization. The pipeline execution model addresses the inefficiencies of blocking operators, while the scalable shuffle tackles the communication bottleneck inherent in shared-nothing architectures. The unified bus and unified remote memory access reduce the cost of data movement, and the cost-based Bloom filter placement with streaming strategies minimizes unnecessary data transfer by filtering early and across nodes. Together, these changes form a holistic approach to improving analytical workload performance at extreme scale. The 40% improvement over the best published result at 30 TB demonstrates the effectiveness of this multi-pronged strategy. Future work may involve further tuning of these components and extending the approach to other workloads and larger scales.

Who should read this

CS practitioners and researchers

Opening member content…