VDTuner: Automated Performance Tuning for Vector Data Management Systems

TL;DR

VDTuner tackles the auto-configuration of VDMS by casting tuning as a multi-objective optimization problem over speed and recall. It introduces a holistic MOBO framework with a polling surrogate, EHVI-based acquisition, and dynamic budget allocation across index types to efficiently discover high-quality configurations without prior VDMS knowledge. Empirical results on Milvus show consistent improvements in search speed and recall, and substantial gains in tuning efficiency compared with strong baselines, including scalability to larger datasets and cost-aware objectives. The work provides a practical, scalable solution for automatic VDMS optimization with broad impact on large-scale IR and ML systems.

Abstract

Vector data management systems (VDMSs) have become an indispensable cornerstone in large-scale information retrieval and machine learning systems like large language models. To enhance the efficiency and flexibility of similarity search, VDMS exposes many tunable index parameters and system parameters for users to specify. However, due to the inherent characteristics of VDMS, automatic performance tuning for VDMS faces several critical challenges, which cannot be well addressed by the existing auto-tuning methods. In this paper, we introduce VDTuner, a learning-based automatic performance tuning framework for VDMS, leveraging multi-objective Bayesian optimization. VDTuner overcomes the challenges associated with VDMS by efficiently exploring a complex multi-dimensional parameter space without requiring any prior knowledge. Moreover, it is able to achieve a good balance between search speed and recall rate, delivering an optimal configuration. Extensive evaluations demonstrate that VDTuner can markedly improve VDMS performance (14.12% in search speed and 186.38% in recall rate) compared with default setting, and is more efficient compared with state-of-the-art baselines (up to 3.57 times faster in terms of tuning time). In addition, VDTuner is scalable to specific user preference and cost-aware optimization objective. VDTuner is available online at https://github.com/tiannuo-yang/VDTuner.

Figures (13)

• Figure 1: Complex configuration space: search speed and recall rate of different system configurations. The red line identifies the high-quality space where the configurations outperform the default setting. The stars mark the configurations that are optimal for both objectives.
• Figure 2: The best index type ($\bullet$) varies with system configs.
• Figure 3: (a, b) The best index type for search speed and recall rate can be very different (other parameters are fixed to their default values). (c) Identifying the most suitable index type demands multiple tuning efforts.
• Figure 4: An illustration of EHVI. In (a), the blue area represents the hypervolume of three Pareto frontier solutions; in (b), the red area represents the EHVI of the newly added solution $\bm{x}_1$, and the green area represents the EHVI of the newly added solution $\bm{x}_2$; $\bm{x}_2$ has higher EHVI than $\bm{x}_1$, which will be considered as a better solution.
• Figure 5: VDTuner's auto-configuration framework: holistically learning a Bayesian Optimization model, but selectively polling one index type at a time.