Enabling K-nearest Neighbor Algorithm Using a Heterogeneous Streaming Library: hStreams

hStreams is a recently proposed (IPDPSW 2016) task-based target-agnostic heterogeneous streaming library that supports task concurrency over heterogeneous platforms. We share our experience of enabling a non-trivial machine learning (ML) algorithm: K-nearest neighbor using hStreams. The K-nearest neighbor (KNN) is a popular algorithm with numerous applications in machine learning, data-mining, computer vision, text processing, scientific computing such as computational biology, astronomy, physics, and others. This is the first example of showcasing hStreams’ ability to enable an ML algorithm. hStreams enabled KNN achieves the best performance achievable by either Xeon © or Xeon Phi 1 by utilizing both platforms simultaneously and selectively.