Hybrid KNN-Join: Parallel Nearest Neighbor Searches Exploiting CPU and GPU Architectural Features

TL;DR

This paper presents HybridKNN-Join, a hybrid CPU/GPU approach for low-dimensional KNN-joins that optimizes workload distribution and maximizes hardware utilization, outperforming pure CPU or GPU methods on large datasets.

Contribution

The paper introduces a novel hybrid CPU/GPU algorithm for KNN-joins that intelligently divides work based on data density, improving performance over existing single-architecture methods.

Findings

HybridKNN-Join outperforms reference implementations on large datasets.

Workload balancing between CPU and GPU enhances overall throughput.

Optimizations increase GPU utilization and reduce load imbalance.

Abstract

K Nearest Neighbor (KNN) joins are used in scientific domains for data analysis, and are building blocks of several well-known algorithms. KNN-joins find the KNN of all points in a dataset. This paper focuses on a hybrid CPU/GPU approach for low-dimensional KNN-joins, where the GPU may not yield substantial performance gains over parallel CPU algorithms. We utilize a work queue that prioritizes computing data points in high density regions on the GPU, and low density regions on the CPU, thereby taking advantage of each architecture's relative strengths. Our approach, HybridKNN-Join, effectively augments a state-of-the-art multi-core CPU algorithm. We propose optimizations that $(i)$ maximize GPU query throughput by assigning the GPU large batches of work; $(ii)$ increase workload granularity to optimize GPU utilization; and, $(iii)$ limit load imbalance between CPU and GPU architectures. We compare HybridKNN-Join to one GPU and two parallel CPU reference implementations. Compared to the reference implementations, we find that the hybrid algorithm performs best on larger workloads (dataset size and K). The methods employed in this paper show promise for the general division of work in other hybrid algorithms.