Scholarlink

In this paper we propose algorithms, inspired by bin packing algorithms, that take into account jobs submitted with deadlines in order to improve the resource utilization of HPC clusters by reducing jobs waiting times. These algorithms have the same advantages as Conservative Backfilling i.e. predictability and non-starvation guarantee. We compared the performance of our more efficient algorithm (Deadline-driven First Fit Backfilling algorithm a.k.a DFFBF) to the popular Conservative Backfilling and EASY algorithms and to the more aggressive modification of EASY that is Fattened Backfilling algorithm. Our evaluation results on 4 different real workloads showed that DFFBF always outperforms these 3 algorithms and the waiting time improvements are up to 57%. We also measured the execution times of our algorithm and proved that DFFBF is fast enough to be used on production platforms. So we recommend using the approach of deadline-driven jobs to manage HPC clusters.

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

On the Optimization of Resources Usage in HPC with a Deadline-Driven Jobs Scheduling Algorithm

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Djobolo Tchimou,
  • Tchimou N’takpé,
  • Djédjé Sylvain Zézé,
  • Gokou Hervé Fabrice Diédié

摘要

In this paper we propose algorithms, inspired by bin packing algorithms, that take into account jobs submitted with deadlines in order to improve the resource utilization of HPC clusters by reducing jobs waiting times. These algorithms have the same advantages as Conservative Backfilling i.e. predictability and non-starvation guarantee. We compared the performance of our more efficient algorithm (Deadline-driven First Fit Backfilling algorithm a.k.a DFFBF) to the popular Conservative Backfilling and EASY algorithms and to the more aggressive modification of EASY that is Fattened Backfilling algorithm. Our evaluation results on 4 different real workloads showed that DFFBF always outperforms these 3 algorithms and the waiting time improvements are up to 57%. We also measured the execution times of our algorithm and proved that DFFBF is fast enough to be used on production platforms. So we recommend using the approach of deadline-driven jobs to manage HPC clusters.