A survey and measurement-based comparison of bandwidth management techniques

Abstract

This article gives a brief tutorial for bandwidth management systems; a survey of the techniques used by eight real-world systems, such as class-based queuing (CBQ), per-flow queuing (PFQ), random early detection (RED), and TCP rate control (TCR); a compact testbed with a set of methodologies to differentiate the employed techniques; and a detailed black-box evaluation. The tutorial describes the needs for the three types of policy rules: class-based bandwidth limitation, session-bandwidth guarantee, and inter/intra-class bandwidth borrowing. The survey portion investigates how the eight chosen commercial/open-source real systems enforce the three policy types. To evaluate the techniques, the designed testbed emulates real-life Internet conditions, such as many simultaneous sessions from different IPs/ports, controllable wide area network (WAN) delay and packet loss rate for each session, and different TCP source implementations. The performance metrics include accuracy of bandwidth management, fairness among sessions, robustness under Internet packet losses and different operating systems, inter/intra-class bandwidth borrowing, and voice over IP (VoIP) quality. The black-box test results demonstrate that (1) only the combination of CBQ+PFQ+TCR can solve the most difficult scenario (multiple sessions competing for the narrow 20kb/s class); (2) the TCR approach may degrade the goodput, fairness, and compatibility even under slight packet loss rates (0.5 percent); (3) without PFQ, TCR and RED have limited ability to isolate the sessions (especially for RED); (4) the G.729 VoIP quality over a 125kb/s access link becomes good only after exercising MSS-clamping to shrink the packet size of the background traffic down to 256 byte.