基于BP神经网络的集群负载预测器

针对由于作业调入调出引起的负载突变,提出了基于通知机制的反传(BP)网络和动态滑动窗口混合预测方法,设计并实现了基于神经网络的负载预测器.该方法在发生突变时,利用动态滑动窗口均值法预测并重新训练样本,训练结束后使用新的BP模型预测.其通知机制能减少预测器的样本识别时间,模型保存机制提供了无需训练样本的机会.测试结果表明,该预测器具有较好的预测精度,能够将大部分预测值的平均误差控制在5%以内,并快速适应突变事件.

BP neural network based predictor of server cluster workload

According to the sudden change of workload caused by job join or exit,a new prediction method combining BP and dynamic sliding window average prediction method have been proposed.A neural network based predictor has been developed.When a job joined or exited,the job scheduler informed the predictor.The predictor employed dynamic sliding window average method to predict,and meanwhile,it retrained the BP model.Upon the completion of retraining,the new BP model was leveraged to predict.The notice mechanism of the combining method could reduce the time of identifying the sudden change workload.In addition,this method provided a model reservation mechanism which offered an opportunity without retraining model when jobs exited.The measurement results showed that the predictor could quickly adapt to the sudden change scenario,and the average percentage error is within 5 %.