数据中心中突发流量管理关键技术研究

数据中心中的突发流量极大影响其性能。随着RDMA等的流行，配合无丢失保障机制PFC的拥塞控制是应对突发流量影响的关键；当链路带宽增长至100Gbps，瓶颈将延伸至服务器I/O，突发流量管理的关键在于调度键值存取操作降低其尾延时。本项目研究这些关键技术：1）利用GIX/M/1模型、概率分布及其Laplace变换刻画突发流量、负载分布及服务器之间的依赖关系，分析键值存储系统中的延时分布。2）改进客户端的副本服务器选择算法，平滑突发流量，控制并发程度，克服反馈信息时效性问题，将键值存取操作恰当的分配给性能和负载时变的副本服务器，降低其尾延时。3）优化服务器端的键值存取操作服务顺序调度算法，克服其未考虑服务器的性能和负载状态变化、副本选择的影响的问题，降低Multiget的完成时间。4）改进配合PFC的拥塞控制算法，在识别PFC引起的拥塞的基础上加快响应速度，缓解PFC触发导致的问题，提升网络性能

The burst traffic in data centers would incur severe performance degradation. With the popularity of RDMA et al., the congestion control mechanism cooperated with the lossless guarantee mechanism PFC becomes the key to resist the negative impacts of the burst traffic. When link capacity increases to 100Gbps, the bottleneck will be at the I/O of computer servers, the key of the burst traffic management is to schedule the key-value access operations such that their tail latency is reduced. In this proposal, we study these key technologies of burst traffic management. 1) Depicting the burst traffic, the load distribution and the dependence among servers with the GIX/M/1 model, the probability distribution and its Laplace transformation respectively, we analyze the latency distribution in the key-value store system. 2) Improving the replica server selection algorithm at the client by smoothing the burst traffic and overcoming the timeliness issue of feedback information, we can properly assign the key-value access operations to different replica servers with time-varying performance and workload, ultimately reducing the tail latency of key-value access operations. 3) Optimizing the scheduling algorithm for the service sequence of key-value access operations at servers by overcoming the problem of ignoring the impacts of the time-varying performance and workload of servers and the replica server selection algorithm, we can further reduce the completion time of Multiget. 4) Improving the congestion control algorithm to cooperate with the PFC mechanism by reducing the response time based on the distinguishing of the congestion generated by PFC, we mitigate the negative impacts incurred by PFC and accordingly improve the network performance in data centers.