Large-scale response-aware online ANN search in dynamic datasets

Abstract

Similarity search is an key operation in Content-based multimedia retrieval (CBMR) applications. Online CBMR applications, which is the focus of this work, have to search in large and dynamic datasets that are updated during the execution while offering low response times. Additionally, these applications are submitted to workloads that vary at runtime. The computing demands in this scenario exceeds the processing power of a single computer, motivating the largescale machines in the domain. Thus, in this work, we proposed a distributed memory parallelization of similarity search that addresses the mentioned challenges. Our solution employs the efficient Inverted File System with Asymmetric Distance Computation algorithm (IVFADC) algorithm as the baseline, which is extended here to support dynamic datasets. Further, we developed a dynamic resource management algorithm, called Multi-Stream Adaptation (MS-ADAPT), that is executed at run-time to change the computing resource assignment with the goal of minimizing response times. We have evaluated our system solution with multiple data partitioning strategies using up to 160 compute nodes and a dataset with 344 billion multimedia descriptors. It demonstrated superlinear scalability for our Spatial-Aware data partition algorithms and MS-ADAPT outperformed the best static approach (oracle) by reducing the response times up to 32× on high-load cases.