Abstract
The data center bridging (DCB) protocols enable Ethernet to become the leading unified fabric for a converged data center. As a part of DCB, priority flow control (PFC) is a control mechanism that provides the losslessness feature required by certain applications such as Fibre Channel over Ethernet. Quantized congestion notification (QCN) is also a DCB protocol and provides congestion control functionality to complement PFC. However, in today's practical data center networks, QCN is not feasible and PFC is not efficient against congestion. We propose a new link layer mechanism called congestion aware priority flow control (CaPFC) that equips PFC with better congestion control capability in the absence of QCN. Changes brought by CaPFC on top of PFC are lightweight. By conducting simulations in NS-3, we demonstrate that CaPFC handles persistent and flash congestion much better than PFC. For the short messaging traffic and the query traffic, which are sensitive to tail latencies, CaPFC consistently outperforms PFC in terms of tail flow completion time.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call