Scholarlink

AI training often need use more than one thousand NPU cards, which adopts the multi-layer CLOS structure. In this paper, we concentrate on increasing the training efficiency for multiple parallel AI training tasks in the CLOS network. First, we use six-dimensional coordinates to represent the positions and transmission paths of each pair of NPUs and analyze CLOS network’s structure. Second, we complete single task optimization by choosing the suitable NPU cluster through heuristic algorithm and avoiding the Traffic Conflict through greedy algorithm. At last, in order to complete parallel multi-task optimization, we decompose and merge tasks into new task sets, and use the same steps as single task optimization to obtain the shortest execution time for the task set.This strategy has the advantages of higher speed and optimization efficiency than before.

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

Multi-layer CLOS Network Task Optimization Based on Heuristic and Greedy Algorithm

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jiaxiang Li,
  • Jun Feng,
  • Yu Wang

摘要

AI training often need use more than one thousand NPU cards, which adopts the multi-layer CLOS structure. In this paper, we concentrate on increasing the training efficiency for multiple parallel AI training tasks in the CLOS network. First, we use six-dimensional coordinates to represent the positions and transmission paths of each pair of NPUs and analyze CLOS network’s structure. Second, we complete single task optimization by choosing the suitable NPU cluster through heuristic algorithm and avoiding the Traffic Conflict through greedy algorithm. At last, in order to complete parallel multi-task optimization, we decompose and merge tasks into new task sets, and use the same steps as single task optimization to obtain the shortest execution time for the task set.This strategy has the advantages of higher speed and optimization efficiency than before.