基于时序相关性的云平台多负载序列联合预测

Joint Prediction of Multi-Workload Sequences Based on Temporal Correlation in the Cloud

为提高云平台负载预测的精度,提出了一种基于时序相关性的多负载序列联合预测方法.首先,为获得相似的负载序列,采用长短时记忆网络提取负载序列的时序特征,再利用层次聚类法,获得在时序特征空间相似的负载序列类;其次,对获得的每个负载序列类分别构建多任务学习模型,挖掘和利用负载序列间隐藏的共享领域知识,提高模型泛化能力和预测精度,并实现多个负载序列的联合预测.使用Google数据集的中央处理器负载监控日志进行验证,结果表明,时序特征聚类可有效提取和利用负载序列的全局时序特征,降低原始序列的噪声,获得特征上相似的序列;与常用的负载预测方法比,所提方法对不同变化规律的负载序列都具有更精确的预测效果.

A novel approach,joint prediction of multi-workload sequences,was proposed based on temporal correlation. Firstly,long short-term memory was used to extract the temporal feature among workload sequences for obtaining the similar workload sequences,while hierarchical clustering algorithm was used to obtain workload sequences clusters. Then,construct multi-task learning model respectively for each obtained sequence clusters,capture and utilize the shared domain knowledge among multiple workload sequences,so as to achieve joint prediction of multiple workload sequences as well as improve generalization ability and prediction accuracy of the model. Results of experiment on dataset of Google cluster trace2011 demonstrates that the temporal feature clustering can effectively extract and utilize the global temporal feature of workload sequences,reduce the noise of original sequences and get workload sequences clusters with similar characteristics. Proposed method performs better in prediction accuracy than the state-ofthe-art methods.