Scholarlink

The article describes the approach on quality evaluation of the distributed HPC (High-performance computing) systems used for random search tasks processing. In such systems, all the tasks approach the system within non-stationary combined flow and have a strict time constraint for their execution. The control of task performance in such systems is conducted on the principles of sporadic control by the intelligent agents. To assess the functioning validity of the distributed HPCs, the concepts of specific intensity of the input flow and integral productivity of the system are introduced in the paper. This allowed us to perform the input (tasks) - output (solutions) analysis for non-stationary input flow. Based on modelling it is shown that Little’s law is fulfilled despite strict time limitations for the fulfillment of each task. We also introduce the concept of resource-time compression coefficient, providing for the comparison of the productivity with the baseline scheduling for the equivalent computational field.

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

Specialized HPC Systems Performance Gained by Discrete Multi-agent Management

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • P. E. Golosov,
  • Sergey Bolovtsov,
  • M. M. Polukoshko,
  • I. M. Gostev

摘要

The article describes the approach on quality evaluation of the distributed HPC (High-performance computing) systems used for random search tasks processing. In such systems, all the tasks approach the system within non-stationary combined flow and have a strict time constraint for their execution. The control of task performance in such systems is conducted on the principles of sporadic control by the intelligent agents. To assess the functioning validity of the distributed HPCs, the concepts of specific intensity of the input flow and integral productivity of the system are introduced in the paper. This allowed us to perform the input (tasks) - output (solutions) analysis for non-stationary input flow. Based on modelling it is shown that Little’s law is fulfilled despite strict time limitations for the fulfillment of each task. We also introduce the concept of resource-time compression coefficient, providing for the comparison of the productivity with the baseline scheduling for the equivalent computational field.