Abstract
Large-scale data centers play a vital role in supporting the ever growing demand for computation and storage. These data centers often employ commodity hardware switches with shallow buffers arranged in a multi-rooted tree topology. Many-to-one traffic communication pattern is often observed in a data center whereby many servers simultaneously send requested data to an aggregation server. This may result in excessive congestion on the bottleneck switch, overwhelming the shallow switch buffer and resulting in large number of packet drops. Legacy transmission control protocol (TCP) treats a packet drop as an indication of severe congestion in the path and forces the sender to drastically reduce its sending rate. This leads to the phenomena of TCP incast resulting in severe throughput collapse deteriorating the performance of the data center. Many solutions have been proposed in literature that addresses the TCP incast issue by either tweaking TCP parameters or proposing a modified/custom version of TCP to be used in data centers. Recently, software-defined networking (SDN) has gained the attention of researchers due to the ease of centralized control and programmability of the network. The centralized nature of SDN presents an opportunity to address the TCP incast issue in data center networks in an efficient manner. This paper provides a snapshot of research efforts that leverage the capabilities of SDN to mitigate the effect of congestion in data centers. We provide a taxonomy of the existing solutions, discuss complementary concepts and highlight their relative strengths along with their shortcomings.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have