Scholarlink

This research investigates the effectiveness of emerging container platforms specifically tailored for high-performance computing (HPC) and scientific applications. Through real-time comparisons between containerized and native environments on the University of Alabama’s HPC infrastructure, we analyze performance overheads and identify their causes in data processing. Utilizing the OSU Micro-Benchmarks suite, MiniVite, and Generative Adversarial Networks (GANs), we evaluate Singularity’s compatibility with HPC requirements. Our experiments involve configuring Singularity containers with UCX and InfiniBand support, utilizing real-world large graph data, benchmarking tests, and testing the GPU utilization during GAN training. We conducted both single-node and multi-node experiments to evaluate performance overheads in different configurations. Our findings demonstrate that the performance overhead in both single and multi-node configurations are minimal in general, with overheads within an acceptable range for HPC compared to native implementations. These results establish containers as a robust platform for contemporary and future scientific computations, offering near-native performance alongside enhanced compatibility, security, and efficiency.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Performance Evaluation of HPC Workloads Through Containerization with Singularity

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Hari Teja Jajula,
  • Purushotham V. Bangalore

摘要

This research investigates the effectiveness of emerging container platforms specifically tailored for high-performance computing (HPC) and scientific applications. Through real-time comparisons between containerized and native environments on the University of Alabama’s HPC infrastructure, we analyze performance overheads and identify their causes in data processing. Utilizing the OSU Micro-Benchmarks suite, MiniVite, and Generative Adversarial Networks (GANs), we evaluate Singularity’s compatibility with HPC requirements. Our experiments involve configuring Singularity containers with UCX and InfiniBand support, utilizing real-world large graph data, benchmarking tests, and testing the GPU utilization during GAN training. We conducted both single-node and multi-node experiments to evaluate performance overheads in different configurations. Our findings demonstrate that the performance overhead in both single and multi-node configurations are minimal in general, with overheads within an acceptable range for HPC compared to native implementations. These results establish containers as a robust platform for contemporary and future scientific computations, offering near-native performance alongside enhanced compatibility, security, and efficiency.