Abstract

The lifecycle of a supercomputer job includes queue waiting time, allocation of computational resources, job initialization, job execution on allocated resources, result saving and resources freeing. The paper focuses on scheduling optimization of supercomputer jobs with a long initialization time, which reduces useful load of the supercomputer. Grouping jobs could increase the useful load of the supercomputer. The paper provides a survey of relevant works in this field, considers various approaches. We propose an approach that groups jobs of different types. The approach was implemented as a job grouping subsystem for domestic job management system SUPPZ, used in JSCC RAS. We provide the results of experiments that confirm effectiveness of the approach proposed.

中文翻译：

---

基于组的作业调度，以提高高性能计算效率

摘要

超级计算机作业的生命周期包括队列等待时间，计算资源分配，作业初始化，在分配的资源上执行作业，节省结果和释放资源。本文着重于优化具有较长初始化时间的超级计算机作业的调度，从而减少了超级计算机的有用负载。分组作业可能会增加超级计算机的有用负载。本文提供了对该领域相关工作的调查，并考虑了各种方法。我们提出了一种对不同类型的工作进行分组的方法。该方法被实现为JSCC RAS中使用的家庭作业管理系统SUPPZ的作业分组子系统。我们提供的实验结果证实了所提出方法的有效性。