Available at: http://digitalcommons.calpoly.edu/theses/602
Date of Award
MS in Computer Science
Hugh M. Smith
Networks within data centers, such as connections between servers and disk arrays, need lossless flow control allowing all packets to move quickly through the network to reach their destination. This paper proposes a new algorithm for congestion control to satisfy the needs of such networks and to answer the question: Is it possible to provide circuit-less reliability and flow control in an Ethernet network? TCP uses an end-to-end congestion control algorithm, which is based on end-to-end round trip time (RTT). Therefore its flow control and error detection/correction approach is dependent on end-to-end RTT. Other approaches utilize specialized data link layer networks such as InfiniBand and Fibre Channel to provide network reliability.
The algorithm proposed in this thesis builds on the ubiquitous Ethernet protocol to provide reliability at the data link layer without the overhead and cost of the specialized networks or the delay induced by TCP’s end-to-end approach. This approach requires modifications to the Ethernet switches to implement a back pressure based flow control algorithm. This back pressure algorithm utilizes a modified version of the Random Early Detection (RED) algorithm to detect congestion.
Our simulation results show that the algorithm can quickly recover from congestion and that the average latency of the network is close to the average latency when no congestion is present. With correct threshold and alpha values, buffer sizes in the network and on the source nodes can be kept small to allow little needed additional hardware to implement the system.