PFO: A Parallel Friendly High Performance System for Online Query and Update of Nearest Neighbors

TL;DR

PFO is a high-performance, parallel-friendly LSH system designed for real-time online nearest neighbor search, offering scalable capacity, reduced latency, and improved neighbor quality for big data applications.

Contribution

It introduces a novel parallel-friendly indexing structure for LSH that handles real-time queries and updates efficiently, scalable with flash memory integration.

Findings

Shorter latency compared to existing LSH systems

Higher throughput in online query/update scenarios

Better neighbor quality in streaming applications

Abstract

Nearest Neighbor(s) search is the fundamental computational primitive to tackle massive dataset. Locality Sensitive Hashing (LSH) has been a bracing tool for Nearest Neighbor(s) search in high dimensional spaces. However, traditional LSH systems cannot be applied in online big data systems to handle a large volume of query/update requests, because most of the systems optimize the query efficiency with the assumption of infrequent updates and missing the parallel-friendly design. As a result, the state-of-the-art LSH systems cannot adapt the system response to the user behavior interactively. In this paper, we propose a new LSH system called PFO. It handles query/update requests in RAM and scales the system capacity by using flash memory. To achieve high streaming data throughput, PFO adopts a parallel-friendly indexing structure while preserving the distance between data points. Further, it accommodates inbound data in real-time and dispatches update requests intelligently to eliminate the cross-threads synchronization. We carried out extensive evaluations with large synthetic and standard benchmark datasets. Results demonstrate that PFO delivers shorter latency and offers scalable capacity compared with the existing LSH systems. PFO serves with higher throughput than the state-of-the-art LSH indexing structure when dealing with online query/update requests to nearest neighbors. Meanwhile, PFO returns neighbors with much better quality, thus being efficient to handle online big data applications, e.g. streaming recommendation system, interactive machine learning systems.