Summary

This paper presents Kubism, a software approach designed to improve the performance and resource utilization of K-means clustering on mobile heterogeneous platforms (e.g., NVIDIA Jetson Orin AGX). Traditional parallel K-means implementations often underutilize GPU warps and leave CPU resources idle. Kubism addresses these inefficiencies by disassembling the K-means algorithm into finer-grained tasks and reassembling them to achieve balanced, dynamic cooperation between CPU and GPU.

Key techniques include:

  • Reordering operations to avoid unnecessary work
  • Balancing workloads across CPU and GPU to reduce idle time
  • Dynamically adjusting task allocation using runtime performance metrics
  • Reconfiguring processing units to minimize waiting and synchronization overhead

Combined, these strategies lead to significant improvements: up to 2.65× speedup in individual clustering iterations and an average 1.23× improvement in overall end-to-end execution time compared to prior work, as measured on a Jetson Orin AGX evaluation platform. The paper details the design, implementation, and empirical evaluation of Kubism.

Download & Original

The full text of the paper is available from the ACM Digital Library and is subject to ACM copyright. This page provides an original English summary and metadata only; consult the ACM PDF for the complete article.