HTC集群与HPC集群负载融合的二阶作业调度算法和资源管理研究

High Energy Physics(HEP) experiments require large scale computing service. To support HEP experiments, two types of clusters are constructed. One is a High Throughput Computing(HTC) cluster, the other is a High Performance Computing(HPC) cluster. Currently the resource utilization ratio of the HTC cluster has reached up to 90% by sharing resources dynamically, however, the problem of the peak resource requirement larger than supplement still exists. On the contrary, the resource utilization ratio of the HPC cluster is only 50% because of lower workload, which means that if the job workload could be integrated between these two clusters, it will not only decrease job queue time of the HTC cluster but also increase the resource utilization ratio of the HPC cluster...To integrate workload between the two clusters, this project will start a research about federated job scheduling algorithms and resource management. To be concrete，the research is consisted with five parts which are a two-staged scheduling algorithm to schedule jobs controllably and effectively, a batch scheduling algorithm to bring transparent workload to the HPC cluster, an intergroup resource sharing system to share resources dynamically, a federated cluster accounting system to provide bills and user query services, and unified user interfaces to access the two clusters with workload integrated. With a series research about job scheduling algorithms, resource manage and supported systems, it is anticipated to decrease job queue time of the HTC cluster by at least 10% and to increase resource utilization ration of the HPC cluster by at least 20%. As a result, it will improve the scheduling efficiency both for the HTC cluster and the HPC cluster, which finally reach the point to provide more computing resource and accelerate the high-quality outcome of the HEP experiments.

高能物理实验需要大规模计算支持。为此，高能所构建了两个计算集群，其一为高通量计算（HTC）集群，另一为高性能计算（HPC）集群。目前，尽管HTC集群通过资源共享的方式将资源利用率提高至90%以上，但仍面临峰值需求大于资源供给的问题。与此相对，HPC集群年平均资源利用率仅为50%。因此，若将HTC集群与HPC集群的负载融合，不仅可减少HTC集群作业排队时间，同时可提高HPC集群的资源利用率。.为了融合作业负载，本项目将进行负载融合作业调度算法和资源管理研究，内容包括五个部分：实现集群间作业可控高效调度的二阶调度算法，实现透明负载的作业批量调度算法，负责资源管理的组间资源共享信息系统，支持负载融合的集群记账系统，以及统一的集群用户接口。通过这些研究，本项目将在不影响现有集群调度效率的基础上，至少减少10%的HTC集群作业排队时间，并将HPC集群的资源利用率至少提高20%，从而加速实验成果产出。

本课题针对高能物理计算HPC集群资源利用率不足、同时HTC集群排队作业量大的问题展开研究，通过研发HTC-HPC负载融合的资源分配和二阶作业调度算法，该算法可根据当前集群状态以及历史作业几率，选择批量HTC作业、并调度到HPC集群按分析算法提供的可用资源槽中，可将HPC集群的整体资源利用率提高20%以上。在HTC-HPC负载融合的资源分配和作业调度过程需要获取组间资源使用权限相关信息，为此，开发了动态的组间资源共享信息系统，为算法提供信息支持。此外，为方便系统运行和用户使用，开发了支持负载融合的作业记账系统以及统一的用户接口，用户可使用统一接口提交并查询HTC-HPC负载融合作业。本课题的研究内容目前已应用在实际生产系统中，不仅包括北京计算中心的HPC和HTC集群之间，还应用于北京计算算中的HTC集群和东莞大科学数据中心的HPC集群，为LHAASO、BESIII等实验提供了更多可扩展计算资源。