Scholarlink

In the distributed virtualized environment, multimodal learning tasks benefit from efficient resource sharing and scheduling to obtain a better quality of service, especially execution time. However, many existing virtualization platforms schedule resources with default configuration, instead of dynamic configuration adjustment based on different resource demands of deployed tasks, severely constraining overall performance. This paper proposes a combined optimization approach for the execution time of deployed multi-modal learning tasks, analyzing each stage of the task in the life cycle model. We significantly reduced the creation time by introducing a mirror volume caching mechanism and dynamic parallel parameters adjustment mechanism. As to the running stage of the task, we use Daemon Process Method to shorten the buffer time. Additionally, to tackle the issue of long scheduling times, the daemon process method we adopted has effectively shortened the scheduling delay. We conducted comparative experiments in OpenStack environment with infrastructures. The results show that our combined optimization approach reduces the total execution time by 7.66%.

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

Acceleration Method of Distributed Multimodal Learning Tasks in Cloud Virtualization Environment

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bo Shan,
  • Xinao Wang,
  • Chang Xu,
  • Shan Jiang,
  • Xianggan Liu,
  • Xuan Liu

摘要

In the distributed virtualized environment, multimodal learning tasks benefit from efficient resource sharing and scheduling to obtain a better quality of service, especially execution time. However, many existing virtualization platforms schedule resources with default configuration, instead of dynamic configuration adjustment based on different resource demands of deployed tasks, severely constraining overall performance. This paper proposes a combined optimization approach for the execution time of deployed multi-modal learning tasks, analyzing each stage of the task in the life cycle model. We significantly reduced the creation time by introducing a mirror volume caching mechanism and dynamic parallel parameters adjustment mechanism. As to the running stage of the task, we use Daemon Process Method to shorten the buffer time. Additionally, to tackle the issue of long scheduling times, the daemon process method we adopted has effectively shortened the scheduling delay. We conducted comparative experiments in OpenStack environment with infrastructures. The results show that our combined optimization approach reduces the total execution time by 7.66%.