Embedded Networks Laboratory Understanding Congestion Control in Multi-hop Wireless Mesh Networks Sumit Rangwala Apoorva Jindal, Ki-Young Jang, Konstantinos Psounis, and Ramesh Govindan
Embedded Networks Laboratory Mesh Networks Static multi-hop mesh networks have been proposed an alternative to wired connectivity User’s satisfaction hinges on transport performance –TCP’s performance on mesh networks is known to be poor Starvation Is poor transport performance inherent to multi-hop mesh networks? Can a correctly designed transport help make mesh networks a viable alternative? 2
Embedded Networks Laboratory TCP’s Performance TCP only signals flows traversing the congested link –Link centric view of congestion Fails to account for neighborhood congestion 3 TCP Optimal (Max Min) What mechanisms can help us achieve near-optimal rates?
Embedded Networks Laboratory WCPCap WCP Approach AIMD Based Design Neighborhood-centric Transport 4 Explicit Rate Notification
Embedded Networks Laboratory Neighborhood of a Link Neighbors (overhearing) 10 Neighborhood of a link –All incoming and outgoing links of Sender Receiver One hop neighbors of the sender One hop neighbors of the receiver Link → sender receiver pair Prohibits channel capture Prohibits channel capture at the sender or causes collision at the receiver Ensuing ACK prohibits channel capture at the sender or causes collision at the receiver
Embedded Networks Laboratory WCP: AIMD Based Design When a link is congested, signal all flows traversing the neighborhood of a link to reduce their rate by half, i.e., r f = r f / 2 React to congestion after RTT neighborhood Multiplicative Decrease Key Insight: Congestion is signaled to all flows traversing neighborhood of a congested link 6
Embedded Networks Laboratory WCP During no congestion increase a flow’s rate as r f = r f + α Every RTT neighborhood Additive Increase Key Insight: Rate adaptation is clocked at the largest flow RTT in a neighborhood RTT neighborhood : Largest flow RTT within the neighborhood 7
Embedded Networks Laboratory Simulations: Stack Topology WCP achieves near optimal performance –Through congestion sharing in the neighborhood Simulation setup –Qualnet –802.11b MAC with default parameters –TCP SACK –Auto rate adaptation is off
Embedded Networks Laboratory WCPCap WCP Approach AIMD Based Design Neighborhood-centric Transport 9 Explicit Rate Notification
Embedded Networks Laboratory WCPCap: Explicit Rate Feedback Estimate residual capacity in a neighborhood –Need to know the achievable rate region for scheduled mesh networks Using only local information 10 Challenge: Is a given set of rates achievable in a neighborhood?
Embedded Networks Laboratory Combine, incorporating link dependencies, individual probabilities to find net collision and idle probabilities of the link Combine, incorporating local link dependencies, individual probabilities to find net collision and idle probabilities for the link Calculating Achievable Rates Decompose the neighborhood topology of a link into canonical two-link topologies Find collision and idle time probability of the link in every two- link topology Compute expected packet service time for a link from collision and idle probability of the link Check feasibility, i.e., for each link, Packet arrival rate × E[service time of a packet] ≤ U, 0 ≤ U ≤ 1 11 Requires global information Using only local information Jindal et. al., “The Achievable Rate Region of Scheduled Multi-hop Networks”.
Embedded Networks Laboratory WCPCap: Explicit Rate Feedback Every epoch –Find, by binary search, the largest increment or smallest decrement, δ, such that the new rates are achievable yet fair –Increase/decrease rate of each flow by δ U=1 (100% utilization) would yield large delays, we target U=0.7 12
Embedded Networks Laboratory Simulations: Stack Topology WCPCap slightly better than WCP –Yields smaller queue and thus smaller delays –Not as good as optimal as we target 70% utilization Simulation setup –Qualnet –802.11b MAC with default parameters –TCP SACK –Auto rate adaptation is off TCP Optimal WCPCap WCP
Embedded Networks Laboratory Simulations: Diamond Topology WCP does not achieve max-min rates –Rates are dependent on the number of congested neighborhood and the degree of congestion WCPCap achieves max-min rates
Embedded Networks Laboratory Experimental Setup Mini-PCs running Click and Linux –ICOP eBox b wireless cards running the madwifi driver Omni directional antennas –some antennas covered with aluminum foils to reduce transmission range 15
Embedded Networks Laboratory Experimental Results: Stack Topology SimulationsExperiments For this topology, WCP’s simulation and experimental results are nearly identical 16
Embedded Networks Laboratory Experimental Results: Arbitrary Topology 14 nodes and five flows TCP starves different flows during different runs WCP consistently gives fair rates 17
Embedded Networks Laboratory Related Work WCP –Congestion control schemes explicitly recognizing neighborhood NRED, EWCCP, and IFRC –Congestion control for ad-hoc wireless networks TCP-F, TCP-ELFN, TCP-BuS, ATCP, etc. COPAS, LRED, ATP, etc. –Congestion control for last-hop wireless networks I-TCP, Snoop, WTCP, etc. WCPCap –Heuristic based capacity estimation WXCP and XCP-b Schemes that also change the MAC layer –e.g, wGDP, DiffQ 18
Embedded Networks Laboratory Conclusions and Future Work Demonstrate plausibility of distributed fair rate control for mesh networks –Low overhead AIMD scheme –Explicit rate feedback scheme Future Work –Optimizing AIMD parameters in WCP –Reduce control overhead of WCPCap –More extensive experiments 19
Embedded Networks Laboratory Thank You 20