拓展的流优先级控制协议设计

Design of Extended Flow Priority Control Protocol

阐述数据中心逐渐采用远程直接内存访问(RDMA)技术来满足高吞吐量(40Gbps)和低延迟(每跳小于10ms)的通信要求。在一个基于IP的数据中心路由网络中,RDMA技术使用RoCEv2(融合以太网的RDMA系统)架构的无损网络。该架构基于队列优先级对数据流进行流量控制(PFC),以确保零数据包丢失。随着发送端口阻塞的出现,PFC机制不断被触发,不可避免地会出现队头阻塞的问题,许多不通过阻塞出端口队列的数据流将被阻塞,极大程度危害无辜流的传输性能。现有的方法采用流控制和拥塞控制的方法,使无害流在拥塞后能迅速恢复到原来的发送速度。在阻塞期间,无辜流量的传输几乎停滞不前。无辜流的流传输完成时间(FCT)随着拥塞时间的增长而线性增长。为了消解这个问题,提出拓展的PFC协议(xPFC,extended PFC),一个基于传统PFC协议的拓展PFC协议。设计应用瞬时和动态的流量优先级转移机制,将受害者流量与拥堵流量解耦。因此,即使在拥堵期间,无辜流的流量也可以通过拥堵的端口。大量的实验模拟验证设计可以有效地缓解受害者流,并且受害者流的FCT增加不到拥堵持续时间的10%。

This paper describes that data centers gradually adopt Remote Direct Memory Access(RDMA) technology to fulfill high-throughput and low-latency communication requirements.In an IPbased routing network,RDMA uses the RoCEv2( RDMA over Converged Ethernet version 2) protocol,which is based on Priority-based Flow Control( PFC) to ensure zero packet loss.However,with the triggering of the PFC mechanism,one of the serious problems called head-of-line blocking inevitably occurs and many innocent flows will be blocked.Existing methods use flow control and congestion control methods,so that the innocent flows can quickly recover the original sending speed after the congestion.However,the transmission of the innocent flows almost stagnate during the blocking period.The flow control time( FCT) of victim flows grows linearly with the congestion duration.To alleviate this problem,this paper introduces the xPFC( extended PFC),an extended PFC protocol based on traditional PFC protocol.Our design applies transient and dynamic flow priority shift mechanism to decouple the victim flows fromthe congested flows.Therefore,the victim flows can pass the congested port even during the congestion period.Our plenty of packet-level simulations verify that our design can effectively alleviate victim flows and the FCT of victim flows increases by less than 10 percent of the congestion duration.