数据中心网络coflow调度机制结构构建及仿真

重新构建得到了一种coflow调度算法-DeepCS,将coflow资源视图看成是需要进行后续处理的图像,根据之前学习策略来达到coflow的最佳调度效果。利用DNN提取特征参数时不必通过人为手动的方法进行设计,通过单独学习过程便可实现,给出深度增强学习系统。训练输入包含了各项网络与任务情景,并以动作概率分布作为输出,EPiSOdE作为单位开展训练过程。仿真结果得到:当coflow到达速率变大后,将会导致所有算法需要更长的coflow完成时间,此时调度算法流时间与的工作压力都会增加,从而形成更长的coflow平均完成时间;在较低的coflow到达速率下,VARYS和DeepCS具有相似的性能,都比PFABRiC的性能更好,并且DeepCS性能提升最快。

Data center network coflow scheduling mechanism structure construction and simulation

DeepCS, a kind of coflow scheduling algorithm, is obtained through reconstruction. The coflow resource view is regarded as the image to be processed later, and the optimal scheduling effect of coflow is achieved according to the previous learning strategy.The feature parameters extracted by DNN need not be designed by manual method, and can be realized by separate learning process.The training input includes various network and task situations, with the motion probability distribution as the output, and EPiSOdE as the unit to carry out the training process.The simulation results are as follows: when the coflow reaches a larger rate, all algorithms will need longer coflow completion time. At this point, the flow time and working pressure of the scheduling algorithm will increase, thus forming a longer average coflow completion time.Under the lower coflow arrival rate, VARYS and DeepCS have similar performance, both of which are better than PFABRiC, and DeepCS has the fastest performance improvement.