GPU-aided edge computing for processing the k nearest-neighbor query on SSD-resident data

Edge computing aims at improving performance by storing and processing data closer to their source. The k Nearest-Neighbor (k-NN) query is a common spatial query in several applications. For example, this query can be used for distance classification of a group of points against a big reference dataset to derive the dominating feature class. Typically, GPU devices have much larger numbers of processing cores than CPUs and faster device memory than main memory accessed by CPUs, thus, providing higher computing power. However, since device and/or main memory may not be able to host an entire reference dataset, the use of secondary storage is inevitable. Solid State Disks (SSDs) could be used for storing such a dataset. In this paper, we propose an architecture of a distributed edge-computing environment where large-scale processing of the k-NN query can be accomplished by executing an efficient algorithm for processing the k-NN query on its (GPU and SSD enabled) edge nodes. We also propose a new algorithm for this purpose, a GPU-based partitioning algorithm for processing the k-NN query on big reference data stored on SSDs. We implement this algorithm in a GPU-enabled edge-computing device, hosting reference data on an SSD. Using synthetic datasets, we present an extensive experimental performance comparison of the new algorithm against two existing ones (working on memory-resident data) proposed by other researchers and two existing ones (working on SSD-resident data) recently proposed by us. The new algorithm excels in all the conducted experiments and outperforms its competitors. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

Edge computing aims to improve performance by storing and processing data closer to their source. This paper proposes a distributed edge-computing environment architecture for large-scale processing of the k-NN query using an efficient algorithm on GPU and SSD-enabled edge nodes. The new algorithm outperforms existing ones in experimental performance evaluations.