Fairness in a data center

Existing data centers utilize several networking technologies in order to handle the performance requirements of different workloads. Maintaining diverse networking technologies increases complexity and is not cost effective. This results in the current trend to converge all traffic into a single networking fabric. Ethernet is both cost-effective and ubiquitous, and as such it has been chosen as the technology of choice for the converged fabric. However, traditional Ethernet does not satisfy the needs of all traffic workloads, for the most part, due to its lossy nature and, therefore, has to be enhanced to allow for full convergence. The resulting technology, Data Center Bridging (DCB), is a new set of standards defined by the IEEE to make Ethernet lossless even in the presence of congestion. As with any new networking technology, it is critical to analyze how the different protocols within DCB interact with each other as well as how each protocol interacts with existing technologies in other layers of the protocol stack.

This dissertation presents two novel schemes that address critical issues in DCB networks: fairness with respect to packet lengths and fairness with respect to flow control and bandwidth utilization. The Deficit Round Robin with Adaptive Weight Control (DRR-AWC) algorithm actively monitors the incoming streams and adjusts the scheduling weights of the outbound port. The algorithm was implemented on a real DCB switch and shown to increase fairness for traffic consisting of mixed-length packets. Targeted Priority-based Flow Control (TPFC) provides a hop-by-hop flow control mechanism that restricts the flow of aggressor streams while allowing victim streams to continue unimpeded. Two variants of the targeting mechanism within TPFC are presented and their performance evaluated through simulation.

Test Service: 
802.1 Ethernet Switching Protocols (ESP) Testing Services
Authors: 
Mikkel Hagen
Category: 
Theses