280 likes | 386 Views
Measurement-based models enable predictable wireless behavior. Ratul Mahajan Microsoft Research Collaborators: Yi Li, Lili Qiu, Charles Reis, Maya Rodrig, Eric Rozner, David Wetherall, John Zahorjan, Yin Zhang, . Wireless Mesh Networks. Can enable ubiquitous and cheap broadband access
E N D
Measurement-based models enable predictable wireless behavior Ratul Mahajan Microsoft Research Collaborators: Yi Li, Lili Qiu, Charles Reis, Maya Rodrig, Eric Rozner, David Wetherall, John Zahorjan, Yin Zhang,
Wireless Mesh Networks Can enable ubiquitous and cheap broadband access Witnessing significant research and deployment But early performance reports are disappointing
Wireless performance is unpredictable • Even basic questions are hard to answer • Arguably the most frustrating aspect of wireless • Mysteriously inconsistent performance • Makes it almost impossible to plan and manage
An example of performance weirdness Bad Good Source Relay Sink Good Bad Source Relay Sink bad-good UDP throughput (Kbps) Simulation bad-good Testbed good-bad bad-good UDP throughput (Kbps) UDP throughput (Kbps) good-bad Loss rate on the bad link 2x good-bad Loss rate on the bad link Source rate (Kbps)
Predictable performance optimization • Given a (multi-hop) wireless network: • Can its performance for a given traffic pattern be predicted? • Can it be systematically optimized per a desired objective such as fairness or throughput? • Yes, and Yes, at least in the context of WiFi
Predictability needs models R1 R2 Success of failure? S1 S2 • To predict if specific nodes interfere and what happens when a set of nodes send together • Without models, we must measure each possibility separately
Traditional wireless models S1 S2 • Typical assumptions • Transmission range is circular • Interference range is twice the transmission range • Then predict the result of various sending configurations
Shortcomings of traditional models • RF propagation is a very complex, esp. indoors • The assumptions almost always do not hold in practice • Great for asymptotic behavior characterization • E.g., expected max throughput as a function of number of nodes • Pretty much useless for predicting behavior in a specific wireless network
A move towards experimentation • Instead of relying on models, test performance of new protocols on testbeds • Hard to say if results generalize • The lack of predictability remains • Unless all possible configurations are tested
Measurement-based models Capture the “RF profile” of the network by measuring simple configurations Use modeling to predict the behavior under more complex configurations Can offer the best of traditional modeling and experimentation worlds
Lessons learned • Simple measurements on off-the-shelf hardware can provide usable RF profile [SIGCOMM2006] • It is possible to model interference, MAC, and traffic in a way that balances fidelity and tractability [MobiCom2007] • Holistically controlling source rates is key to achieving desired outcomes [HotNets2007, SIGCOMM2008]
Measurement-based modeling and optimization Measure the RF profile of the network Constraints on sending rate and loss rate of each link Find compliant source rates that meet the objective
Measurements One or two nodes broadcast at a time • O(n2) measurements Other nodes listen and log received packets Yields information on loss and carrier sense probabilities Measure the RF profile of the network Constraints on sending rate and loss rate of each link R S1 S2 Find compliant source rates that meet the objective
Modeling Measure the RF profile of the network • Makes no assumptions about topology, traffic, or MAC • Lightweight yet realistic • O(# active links) constraints capture the feasible operating region • Throughput constraints • Loss rate constraints • Sending rate constraints Constraints on sending rate and loss rate of each link Find compliant source rates that meet the objective
Throughput constraints • Divide time into variable-length slot (VLS) • 3 types of slots: idle, transmission, deferral Expected payload transmission time Probability of starting transmission in a slot Success probability Expected slot duration
Loss rate constraints • Inherent and collision loss are independent • Inherent loss is directly measured • Collision loss • Synchronous loss • Two senders can carrier sense each other • Occur when two transmissions start at the same time • Asynchronous loss • At least one sender cannot carrier sense the other • Occur when two transmissions overlap
802.11 unicast Random backoff interval uniformly chosen [0,CW] CW doubles after a failed transmission until CWmax, and restores toCWmin after a successful transmission Sending rate feasibility constraints SIFS ACKTransmission DIFS Random Backoff Data Transmission Expected contention window size under loss rate pi
RTS/CTS Add RTS and CTS delay to VLS duration Add RTS and CTS related loss to loss rate constraints Multi-hop traffic demands Link load routing matrix e2e demand Routing matrix gives the fraction of each e2e demand that traverses each link TCP traffic Update the routing matrix: where reflects the size & frequency of TCP ACKs Extensions to the basic model
Optimization Inputs: • Traffic matrix • Routing matrix • Optimization objective • Total throughput, fairness, … Output: • Per-flow source rate Predictable:output rates are actually achievable Measure the RF profile of the network Constraints on sending rate and loss rate of each link Find compliant source rates that meet the objective
Flow throughput feasibility testing • Building block for optimization • Uses an iterative procedure Input: throughput Output:feasible/infeasible Initialize τ= 0 and p = pinherent Estimate τ from throughput and p no Converged? Check feasibility constraints yes Estimate p from throughput andτ
Fair rate allocation Initialization: add all demands to unsatSet Scale up all demands in unsatSetuntil some demand is saturated or scale1 yes if (scale 1) no Move saturated demands from unsatSet to X yes If unsatSet≠ no Output X
Formulate a non-linear optimization problem (NLP) Solve NLP using iterative linear programming Total throughput maximization Maximize total txput Link load is bounded bythroughput constraints Sending rate is feasible E2e throughput is bounded by demand
The network is capable of achieving its model-predicted throughput UDP TCP Results for a 19-node testbed
The network cannot achieve higher than model-predicted throughput UDP TCP
Measurement-based models enable fair throughput distribution (predictably) UDP TCP
Measurement-based models boost network throughput (predictably) TCP UDP
Future work: Making it real Online measurement of RF profile Decentralized computation of source rates Joint optimization of routing and source rates
Conclusions • Wireless behavior is unpredictable • Complex RF propagation • Interactions between MAC, traffic, and interference • Measurement-based models: a new approach to obtain predictable behavior • Measure the RF profile and model the rest • Promising results in our experiments on real test beds • Enables predictable optimization