MMDBench: A Benchmark for Hybrid Query in Multimodal Database

Multimodal data, integrating various types of data like images, text, audio, and video, has become prevalent in the era of big data. However, there is a gap in benchmarking specifically designed for multimodal data, as existing benchmarks primarily focus on traditional and multimodel databases, lacking a comprehensive framework for evaluating systems handling multimodal data. In this paper, we present a novel benchmark program, named MMDBench, specifically designed to evaluate the performance of multimodal databases that accommodate various data modalities, including structured data, images, and text. The workload of MMDBench is composed of eleven tasks, inspired by real-world scenarios in social networks, where multiple data modalities are involved. Each task simulates a specific scenario that necessitates the integration of at least two distinct data modalities. To demonstrate the effectiveness of MMDBench, we have developed a hybrid database system to execute the workload and have uncovered diverse characteristics of multimodal databases in the execution of hybrid queries.