Analysis Simulation and Implementation of VCP-Wireless Profiling
Every congestion control protocol operating in wireless networks is potentially faced with two major challenges of performance degradation. These sources are (a) the coupling of fairness and efficiency, and (b) not properly differentiating between congestion-caused loss associated with network buffering and error-caused loss associated with fading effects. In this paper, we provide a VCP-based cross-layer framework of congestion control that can address both challenges noted above. As a part of our framework, we introduce a loss differentiation heuristic algorithm that can be used with a variety of congestion control protocols. Then using analysis, simulation, implementation, and emulation, we profile the performance of a number of congestion control alternatives in wireless networks. We describe the first implementation of VCP as a collection of loadable kernel modules along with fine-tuned implementations of XCP and TCP/AQM+ECN in Linux. We utilize NS2 as our simulation tool and a wired Linux testbed emulating wireless link effects as our experimental tool. We implement a finite-state Markov chain in both NS2 and our testbed in order to model error-caused loss over wireless links. We further use link layer FEC codes on a per packet basis to compensate against such loss. Our profiling results demonstrate that VCP equipped with our loss differentiation heuristic and link layer FEC represents a well-performing yet practical alternative of wireless congestion control. We also identify some of the shortcomings of VCP including its oscillatory behavior in the presence of link estimation errors and poor fairness characteristic in multi-bottleneck networks.
A S surveyed by [1], the coupling of fairness and efficiency is a known source of performance degradation for TCP and end-to-end TCP-based Active Queue Management (AQM) schemes. The problem is further pronounced in networks including high Bandwidth-Delay Product (BDP) links such as wireless and satellite networks. In the past few years, a wide variety of techniques have emerged to increase the efficiency of congestion control protocols in high BDP networks. The works of adaptively adjust the sending window size of TCP by amending the parameters of Additive-Increase Multiplicative-Decrease (AIMD) [10], while the works of use different congestion signals and the works of explicitly signal the congestion information to the sender. Since all of the proposed Parts of this paper appeared in the Proc. of research WCNC’06 and research MILCOM’07. The authors are with the Department of EECS, University of California, Irvine. e-mail: ( [xiaolonl,hyousefi]@uci.edu). This work was sponsored by grant no. COM06-10223 from the Boeing Company and UC Discovery Industry-University Cooperative Research Program. works above retain an integrated controller design, they often fail to achieve both efficiency and fairness. In contrast, recently proposed eXplicit Congestion-control Protocol (XCP) [17] and Variable-structure Congestion-control Protocol (VCP) [18] attempt at decoupling fairness from efficiency in order to address performance degradation of TCP. While XCP uses Multiplicative-Increase MultiplicativeDecrease (MIMD) for efficiency and AIMD for fairness control, VCP applies Multiplicative-Increase, Additive-Increase, and Multiplicative-Decrease (MIAIMD) policies in three regions of congestion known as low-load, high-load, overload regions, respectively. By encapsulating congestion related information into packet headers, both protocols exhibit high utilization and great fairness characteristics while maintaining low persistent queue lengths and reducing congestion-caused loss in wired networks. However, XCP calls for using multiple bits in a packet header introducing significant deployment obstacles. To the contrary, VCP is able to achieve a performance comparable to XCP using only two ECN bits while keeping compatibility with a variety of existing protocols. Transmission over wireless links is subject to both error- and congestion-caused loss. Thus, the performance of any protocol utilizing IP header bits to carry congestion information may be seriously crippled in wireless networks without proper compensation against error-caused loss. While a spectrum of research works as reviewed by [19] have studied congestion control in wireless networks, they often focus on capturing the effects of data link layer in the performance of congestion control. For example, [20] shows that retransmission at the link layer may result in delivering feedback that is misleading to TCP.