Measurement-based TCP parameter tuning in cloud data centers

TCP congestion control has been a native part of all modern Operating System implementations where parameters are initialized assuming an underlying high Bandwidth Delay Product (BDP) environment. However, the significantly lower BDP in Data Centre (DC) networks makes such conservative transport-layer parameters together with deep-buffered switches and bursty traffic a factor of performance degradation, eventually leading to throughput incast collapse. In this paper, we propose a Software Defined Networking (SDN) approach to tune TCP initial window and retransmission timers for newly created flows based on a network-wide view created by aggregating known characteristics and temporal measurements at a central controller. Through simulation, we show the detrimental effect static TCP parameters have on mice flows and demonstrate the benefits of network-aware per-flow tuning. We show that the average latency under bursty traffic can be improved by a factor of eight, and that flow start and completion times can be improved by a factor of two and five, respectively.