面向应用定义优先级调度的用户态协议栈研究

User-space network stack with application-defined priority scheduling

针对数据中心负载中混杂请求对延迟敏感型请求响应尾延迟产生干扰,而现有研究无法在不同负载场景下为延迟敏感型请求提供灵活优先调度的问题,本文提出了应用定义优先级调度的用户态协议栈。该设计利用用户态协议栈在数据中心请求处理所处的关键位置,支持上层应用根据负载特征灵活定制优先级识别与调度策略,并由协议栈为优先级识别提供数据包、传输控制协议(TCP)流等丰富的状态信息,实现了不需要改动网络协议栈,就可以对不同负载场景实现灵活的应用定义优先级识别与调度,从而避免延迟敏感型请求受到其他混杂负载请求带来的排队延迟与阻塞延迟干扰。实验结果表明,在不同的负载场景下,通过灵活准确的应用定义优先级调度,可以将延迟敏感型请求的响应尾延迟降低98.5%,有效保障了用户体验。

The mixed requests in the data center introduce the problem of the interference to the tail response latency of latency-sensitive requests response.However,existing researches cannot provide flexible priority scheduling for latency-sensitive requests under different load scenarios.A user-space network stack that applies application-defined priority scheduling is proposed.This design takes the advantage of the key position of the user-space network stack in the data center request processing,supports upper-layer applications to flexibly customize priority identification and scheduling strategies according to load characteristics,and provides rich status information such as packets and TCP flows for priority recognition,so as to realize flexible application-defined priority identification and scheduling for different load scenarios without changing the network stack.This work prevents latency-sensitive requests from being disturbed by queuing and blocking delays caused by the other promiscuous request load.Experimental results show that under different load scenarios,flexible and accurate application-defined priority scheduling can reduce the response tail latency of latency-sensitive requests by 98.5%,which effectively ensures the user experience.