基于瓶颈感知的多级反馈队列Coflow调度机制

Coflow作为并行计算框架的典型流量模型，降低Coflow的完成时间（CCT）成为云计算领域的研究热点。现有Coflow调度机制未考虑云数据中心内网络瓶颈问题，容易造成网络拥塞，导致CCT增加。针对该问题，构建基于瓶颈感知的Coflow调度机制Bamq。利用Lagrange对偶优化Coflow调度模型，以加快Coflow流速并增大吞吐量，从而降低CCT。通过设计多级反馈队列机制，降低吞吐量对网络拥塞产生的影响，根据已发流的大小、宽度和流速信息，构建瓶颈因子以动态调整多级队列的优先级，实现拥塞感知，提高Coflow调度性能。在Facebook真实数据集上进行实验，结果表明，相比Baraat、Varys、Aalo机制，该机制的CCT平均缩短21.3%，吞吐量平均提高17.9%，能够有效提高链路的利用率。

Coflow Scheduling Mechanism for Multi-Level Feedback Queue Based on Bottleneck Perception

Coflow is a typical traffic model based on a parallel computing framework.Reducing the Coflow Completion Time (CCT) has become a popular research topic in cloud computing.The existing Coflow scheduling mechanism does not consider the network bottleneck in the cloud data center, which easily causes network congestion and increases the CCT.Hence, Coflow scheduling mechanism, Bamq, based on bottleneck perception is constructed.The Lagrange duality is used to optimize the Coflow scheduling model such that the Coflow flow rate and throughput are increased, whereas the CCT is reduced.The multi-level feedback queue mechanism reduces the effect of throughput on network congestion.Based on the size, width, and flow rate of the sent flow, the bottleneck factor is constructed to dynamically adjust the priority of multi-level queue, realize congestion perception, and enhance the performance of Coflow scheduling.Experiments on the Facebook dataset show that compared with the Baraat, Varys and Aalo mechanisms, the CCT of this mechanism is reduced by 21.3%, whereas the throughput is increased by 17.9% on average.The proposed mechanism can significantly reduce the CCT and effectively improve link utilization.