DBSCAN算法在高性能计算中心用户分类的应用研究

为提高集群资源使用效率, 管理员需要对用户进行分类, 从而对不同用户提出资源使用策略。DBSCAN(Density Based Spatial Clustering of Applications with Noise)聚类算法可对用户进行分类, 但对初始参数敏感。为此, 提出改进算法, 首先将密度进行层次划分, 由此得出各层次的密度阈值, 在每种阈值下采用DBSCAN算法, 解决全局参数问题。在此基础上, 创新地使用一个直接可达距离排序队列, 将排序信息作为可变参数, 减小初始参数对结果的影响。通过高性能计算中心用户数据的实例验证了其可行性。实验结果表明, 改进后的算法提高了用户分类的准确性和全面性。

Application Research of DBSCAN Algorithm Based on High-Performance Computing Center Users Classification

To enhance service efficiency on cluster resource, administrator needs to make classification of users, and provide various strategies on resource utilization to different users. DBSCAN (Density Based Spatial Clustering of Applications with Noise) algorithm can achieve usersclassification, but the initial parameters are very sensitive. The improved algorithm classifies the level of density firstly, then gets the densitythreshold of each level, and uses DBSCAN under each threshold which solves the problem of global parameters. It uses a sorted queue of directly accessible distance as an innovation, makes the sorting information as variable parameter to decrease the influence of initial parameter. The algorithm has verified its feasibility through example data of HPC users. The experimental result demonstrates that this improved algorithm can achieve a more accurate and comprehensive user classification.