高效支持负载聚合的程序资源敏感度获取及分析方法

发布时间：2018-06-05 09:36

本文选题：程序资源敏感度 + 负载聚合　；参考：《西安交通大学学报》2017年04期

【摘要】：针对集群或数据中心中负载聚合导致的程序和系统性能不稳定的问题,提出一种程序资源敏感度获取及分析方法。该方法利用Linux中的Cgroups机制设置可调整资源大小的控制组,并让程序在控制组内执行,获得程序在资源受限时的性能;将程序在多个资源限制下的敏感度表示成敏感度超平面,从而预测程序协同执行时的性能;将程序所需的资源按照重要程度排序以帮助集群调度。实验结果表明,该方法预测出的敏感度与真实值的平均绝对误差为0.093。与最小负载的簇级调度策略相比,利用程序资源敏感度信息的调度策略可以提升程序性能22.2%,还可以提升磁盘读写带宽利用率和网络带宽的利用率,程序资源敏感度能够高效地支持负载聚合。
[Abstract]:Aiming at the instability of program and system performance caused by load aggregation in cluster or data center, a method of program resource sensitivity acquisition and analysis is proposed. The method uses the Cgroups mechanism in Linux to set up a control group that can adjust the resource size, and allows the program to execute in the control group to obtain the performance of the program when the resource is limited, and the sensitivity of the program under multiple resource constraints is expressed as a sensitivity hyperplane. In order to predict the performance of the program when it is executed in cooperation, the resources needed by the program are sorted according to the degree of importance to help the cluster scheduling. The experimental results show that the average absolute error between the predicted sensitivity and the real value is 0.093. Compared with the cluster-level scheduling strategy with minimum load, the scheduling strategy using program resource sensitivity information can improve the program performance 22.2%, and can also improve the utilization of disk read / write bandwidth and network bandwidth. Program resource sensitivity can efficiently support load aggregation.
【作者单位】：北京航空航天大学计算机学院;中山大学数据科学与计算机学院;
【基金】：国家重点基础研究发展计划资助项目(2016YFB1000503) 国家自然科学基金资助项目(61133004,61361126011,61502019)
【分类号】：TP368.5

【参考文献】