Abstract: Driven by the need of prevailing Web applications and services (e.g., search, online retailing, and cloud computing), data centers (DCs) have been built at an unforeseen rate and scale around the globe in the recent decades. In particular, data center networks (DCNs) have drawn great attention from both academia and industry. Under such a background, this paper surveys one of the key aspects of DCN—transport layer protocol. While transport protocol has a long history on the Internet, it is seldom systematically explored in the DCN context until 2010s. DCN presents different characteristics (e.g., single administrative domain, and homogeneous network structure) from the Internet. This brings about both opportunities and challenges for transport protocol design over it. Motivated by this, a bunch of transport protocols have been proposed. This paper classifies the early work (2010—2015) on DCN transport design into three categories—endhost-based congestion control, switch-assisted arbitration, and network priority scheduling. The paper discusses the pros and cons of the work from the three categories. At last, the paper analyzes the recent research trend on DCN transport design—receiver-driven proactive congestion control, and RDMA (remote direct memory access) transport design.

Key words: data center network, transport protocol, congestion control, explicit congestion notification (ECN), remote direct memory access (RDMA)