Parallelization Strategies for Graph-Code-Based Similarity Search

Abstract

The volume of multimedia assets in collections is growing exponentially, and the retrieval of information is becoming more complex. The indexing and retrieval of multimedia content is generally implemented by employing feature graphs. Feature graphs contain semantic information on multimedia assets. Machine learning can produce detailed semantic information on multimedia assets, reflected in a high volume of nodes and edges in the feature graphs. While increasing the effectiveness of the information retrieval results, the high level of detail and also the growing collections increase the processing time. Addressing this problem, Multimedia Feature Graphs (MMFGs) and Graph Codes (GCs) have been proven to be fast and effective structures for information retrieval. However, the huge volume of data requires more processing time. As Graph Code algorithms were designed to be parallelizable, different paths of parallelization can be employed to prove or evaluate the scalability options of Graph Code processing. These include horizontal and vertical scaling with the use of Graphic Processing Units (GPUs), Multicore Central Processing Units (CPUs), and distributed computing. In this paper, we show how different parallelization strategies based on Graph Codes can be combined to provide a significant improvement in efficiency. Our modeling work shows excellent scalability with a theoretical speedup of 16,711 on a top-of-the-line Nvidia H100 GPU with 16,896 cores. Our experiments with a mediocre GPU show that a speedup of 225 can be achieved and give credence to the theoretical speedup. Thus, Graph Codes provide fast and effective multimedia indexing and retrieval, even in billion-scale use cases.