首页|期刊导航|计算机与现代化|支持混合业务的异构算力调度平台设计与实现

支持混合业务的异构算力调度平台设计与实现OA

Design and Implementation of a Heterogeneous Computing Scheduling Platform Supporting Hybrid Services

中文摘要英文摘要

随着各领域对大规模计算能力需求的急剧攀升,资源调度复杂度因资源规模扩张及架构异质性等因素而显著增加.针对这一挑战,本文设计并实现一种支持混合业务执行的异构算力调度平台(HCSP),通过统一业务模型与资源调度框架,实现对多元化资源的协同调度,有效提升资源利用率和调度效率.设计DAG统一业务模型,以适应混合的业务场景,保证不同业务场景在统一调度框架下混用资源;设计统一调度框架,集成多维度资源调度策略,优化异构资源配置,为资源调度算法提供了基础;引入混合业务到HCSP,支持大规模异构算力环境的高并发调度,提升资源使用效率.实证表明,HCSP在大规模异构算力集群中,显著提升资源利用率,CPU平均利用率由18%跃升至36%,相对Jenkins调度,调度效率有较大提升.该平台不仅展示了其作为先进调度平台的实用性,也为解决大规模计算资源管理难题提供了创新方案.

With the rapid increase of large-scale computing requirements in each field,the complexity of resource scheduling is greatly increased due to resource scale expansion and architecture heterogeneity.To address this challenge,this document de-signs and implements a heterogeneous computing power scheduling platform(HCSP)that supports mixed service execution.Through a unified service model and resource scheduling framework,this platform achieves collaborative scheduling of diversi-fied resources,effectively improving resource utilization and scheduling efficiency.The DAG unified service model is designed to adapt to mixed service scenarios and ensure that resources are mixed in different service scenarios under the unified scheduling framework.Designs a unified scheduling framework,integrates multi-dimensional resource scheduling policies,optimizes het-erogeneous resource configurations,and provides the basis for resource scheduling algorithms.Hybrid services are introduced to HCSP to support high-concurrency scheduling in large-scale heterogeneous computing environments and improve resource usage efficiency.Empirical evidence shows that in large-scale heterogeneous computing-power clusters,HCSP significantly improves resource usage,and the average CPU usage is increased from 18%to 36%.Compared with Jenkins scheduling,HCSP greatly im-proves scheduling efficiency.This platform not only shows its practicality as an advanced scheduling platform,but also provides innovative solutions to the management problems of large-scale computing resources.

胡继东;马欣;韩冰;王天齐;鞠炜刚;朱政;汪鹏

北京兴云数科技术有限公司,北京 100176北京兴云数科技术有限公司,北京 100176北京兴云数科技术有限公司,北京 100176北京兴云数科技术有限公司,北京 100176北京兴云数科技术有限公司,北京 100176北京兴云数科技术有限公司,北京 100176东南大学,江苏 南京 211189

信息技术与安全科学

异构算力任务编排任务调度资源调度调度算法

heterogeneous computing powertask orchestrationtask schedulingresource schedulingiesscheduling algorithm

《计算机与现代化》 2026 (1)

91-100,10

国家自然科学基金资助项目(62376057)

10.3969/j.issn.1006-2475.2026.01.013

评论