A Coding-based Approach to Mitigate TCP Incast in Data Center Networks

摘要

As the key infrastructure of cloud computing, data center network provides routing and transport services for cloud applications. To ensure reliable data delivery, TCP has been used in data centers as the de facto transport layer protocol. However, recently researchers found that when a client has a barrier-synchronized requirement and requests data from many servers, the good put may experience collapse. This phenomenon is well-known as TCP In cast. The main cause of TCP In cast is the scare buffer size in low end switches as well as the mismatch between RTO and RTT in data centers. So existing solutions focus on either minimizing the effect of TCP's RTO, or preventing flows from aggressively utilizing the switch buffer. Different from the previous approaches, in this paper we employ coding-based UDP to replace TCP as the transport protocol for the many-to one applications. Specifically, we use LT Codes to ensure reliable data delivery, and adopt TCP Friendly Rate Control (TFRC) to achieve congestion control. The NS based simulation shows that our approach can significantly enhance the application good put in a typical TCP In cast scenario.