Approximate Similarity Search for Online Multimedia Services on Distributed CPU-GPU Platforms

Abstract: Similarity search in high-dimentional spaces is a pivotal operation found a variety of database applications. Recently, there has been an increase interest in similarity search for online content-based multimedia services. Those services, however, introduce new challenges with respect to the very large volumes of data that have to be indexed/searched, and the need to minimize response times observed by the end-users. Additionally, those users dynamically interact with the systems creating fluctuating query request rates, requiring the search algorithm to adapt in order to better utilize the underline hardware to reduce response times. In order to address these challenges, we introduce hypercurves, a flexible framework for answering approximate k-nearest neighbor (kNN) queries for very large multimedia databases, aiming at online content-based multimedia services. Hypercurves executes on hybrid CPU--GPU environments, and is able to employ those devices cooperatively to support massive query request rates. In order to keep the response times optimal as the request rates vary, it employs a novel dynamic scheduler to partition the work between CPU and GPU. Hypercurves was throughly evaluated using a large database of multimedia descriptors. Its cooperative CPU--GPU execution achieved performance improvements of up to 30x when compared to the single CPU-core version. The dynamic work partition mechanism reduces the observed query response times in about 50% when compared to the best static CPU--GPU task partition configuration. In addition, Hypercurves achieves superlinear scalability in distributed (multi-node) executions, while keeping a high guarantee of equivalence with its sequential version --- thanks to the proof of probabilistic equivalence, which supported its aggressive parallelization design.

• The paper introduces Hypercurves, a novel approach that extends multidimensional indexing into a parallel CPU–GPU framework for approximate k-nearest neighbor searches.
• It employs a dynamic scheduling algorithm that responsively partitions tasks between CPUs and GPUs, achieving up to 30× speedup and superlinear scalability.
• The framework maintains high result quality through probabilistic equivalence while efficiently managing fluctuating query rates in large-scale multimedia services.

Overview of Approximate Similarity Search for Online Multimedia Services on Distributed CPU–GPU Platforms

The paper introduces Hypercurves, an innovative approach for performing approximate similarity search within online multimedia services utilizing distributed CPU-GPU platforms. The research addresses the significant challenges posed by the volume of data and the dynamic interaction of users with the system, leading to fluctuating query rates. Hypercurves sets itself apart by executing on hybrid CPU-GPU environments and employing a novel dynamic scheduler that maximizes resource utilization by cooperatively leveraging the capabilities of both CPUs and GPUs to manage substantial query request rates.

Key Contributions

The paper's key contributions are outlined as follows:

1. Multicurves Foundation: Hypercurves builds on the sequential multidimensional index Multicurves and transitions it into a parallel framework through Anthill, a dataflow-based system. Multicurves itself addresses the challenges of executing approximate k-nearest neighbor (kNN) searches, which are foundational tasks in high-dimensional similarity searches across multimedia databases.
2. Dynamic Scheduling: A notable feature of Hypercurves is its dynamic scheduling algorithm, which dynamically partitions work between CPU and GPU. The scheduler is responsive to variations in query request rates, allowing better utilization of system resources and reduction in query response times.
3. Heterogeneous Environment Optimization: The implementation on heterogeneous platforms showcases performance improvements, including up to 30× speedup against a single CPU-core implementation. This is achieved through optimizations that balance the load between CPU and GPU, leveraging a double-buffering scheme that reduces idle time on these processors.
4. Probabilistic Equivalence: A highlight of the implementation is probabilistic equivalence. The algorithm guarantees that the quality of search results is equivalent to its sequential implementation with high probability, allowing for aggressive parallel optimization without loss of result quality.
5. Superlinear Scalability: Hypercurves demonstrates superlinear scalability in distributed environments as a direct consequence of its efficient workload distribution method, which is crucial for handling modern large-scale databases used in multimedia services.

Implications and Future Directions

The adoption of Hypercurves presents significant implications:

• Practical Deployments: By providing a robust mechanism for handling fluctuating loads efficiently, it becomes highly suited for real-world applications where response time is critical, such as web-based image searches, song identification, and mobile video indexing.
• Scalability and Efficiency: The paper supports the growth of large-scale services by demonstrating scalability and efficient utilization of commodity hardware. The integration of GPU computation underlines its fit for power-efficient large-scale computations.
• Framework Extensions: By showcasing CPU-GPU collaboration within the Anthill framework, Hypercurves opens avenues for future development in handling dynamic, distributed multimedia search tasks.

This research moves the field forward by addressing core challenges associated with similarity searches in multimedia databases, leveraging modern hardware capabilities, and offering a scalable solution adaptable to the dynamic nature of web-based services. Future research could explore further refinements in the scheduling algorithms or investigate the application of these techniques in related areas, potentially extending beyond multimedia retrieval into broader domains of distributed computing.