Emergent Structure Among SelfOrganizing Devices M Steenstrup Stow

Emergent Structure Among Self-Organizing Devices M. Steenstrup Stow Research L. L. C. steenie@rcn. com This work was funded in part by DARPA and ONR.

Pervasive Networking User expectations: - access communications services - independent of time, location, and type of service - without intimate knowledge of algorithms and resources for providing services Challenges of realization: - mix of preexisting infrastructure and spontaneous structure - heterogeneous, selfish devices - operating in time-varying, unpredictable, and potentially hostile environment - communications must be consistent with handling needs of traffic, capabilities of devices, and policies for resource use

Environmental Challenges Wireless communications: - distance-based attenuation - interference - shadowing - multipath fading - broadcast Attacks: - detection - interception - spoofing - replays - denial of service

Heterogeneity of Devices Function: - sensors - actuators - computers - switches Mobility: - frequency - speed - trajectory - autonomy Capacity of resources: - transmission - power supply - computation - storage Algorithms controlling use of resources: - dependent on medium - new and legacy Requirement: sufficient commonality to enable end-to-end communication

Device Behavior Adaptive: - goal-driven - cognizant of current state - state-dependent action to achieve goals Opportunistic: - takes advantage of available information and resources Selfish: - committing resources - defending against attacks Stingy: - reluctant to share resources - expects compensation for services - ignores those that fail to reciprocate Requirement: sufficient cooperation to enable end-to-end communication

Controllable Aspects of the Network spectrum use channel between two nodes network graphs for routing temporal and spatial traffic loads and patterns services provided to applications energy dissipation and dispersion vulnerability to attacks that degrade or deny service

The Problem of Network Control Objectives State Constraints Control Algorithm Network Environment Actions

Fundamental Differences Among Controllers Objectives State Constraints Control Algorithm Network Environment Actions goals actions inputs open or closed loop responsiveness duration of response extent of response effects on others

Degraded Signals Objectives State delay noise loss Constraints Control Algorithm Network Environment Actions delay noise loss

Network Control Algorithms Behavioral objectives: - autonomous execution by devices - synergistic, not destructive, interactions with other controllers - graceful degradation with quality of state and action signals - minimal use of resources Design rules of thumb: - aim for improved performance - inject some randomness into selected actions to help prevent unwanted synchronization and to explore space of possible outcomes - primary greedy algorithm based on local state information - secondary cleanup algorithm employed when result from primary algorithm deemed unacceptable

Self-Organizing Devices Two network control problems: - negotiating frequencies for common channels among nodes - selecting next hops for packet forwarding Corresponding emergent structures: - clusters of nodes sharing common channels - paths through network from sources to destinations Common features: - start with tabula rasa - structure created with distributed algorithm - no one node has global knowledge of structure - acceptable performance with respect to end-to-end service is achieved with only local knowledge of structure

Selecting Frequencies for Common Channels Dependent upon: - supported frequencies - spectrum use policies of primary users - measured characteristics of channel - current environment - current location - type of traffic Algorithms: - channel assertion, consolidation, and negotiation - diffusion-based labelling according to constraints - number of channels: 1 best-case, N worst-case - number of transmissions: N best-case, N 2 worst-case

Channel Detection and Selection Assumptions for each node: - set of f supportable frequencies - spectrum use policies for primary users - information communicated as signatures Measure spectrum: Repeat r times initially and whenever possible thereafter Select at random ordering of f frequencies Sample each frequency for t seconds Classify sampled frequencies: - frequency unavailable if active primary user detected in band or spectrum use policy precludes use - available frequency busy if measured use exceeds threshold - frequency removed from list of active channels if used by any node for more than given time interval

Channel Detection and Selection Detect existing channels: When signature detected Record frequency in list of channels If frequency available and no channel yet adopted Suspend sampling temporarily Adopt frequency and assert channel Resume sampling Establish new channel: If no signature detected on available frequency during sampling Select least-busy available frequency Record frequency in list of channels Adopt frequency and assert channel

Channel Detection and Selection Assert channel: If asserting channel in response to detected signature Select at random one of s silent beaconing intervals Repeat 2 ft times Broadcast signature on adopted frequency during selected interval Remain silent for next s beaconing intervals

Assigning Initial Channels

Clusters of Common Channels

Channel Consolidation Assumptions for each node: - initial spectrum sampling concluded - adopted frequency - list of frequencies of all channels overheard - information communicated as packets - notion of common time Initiate consolidation: If g > 1 available frequencies in channel list If adopted frequency higher than others in list Generate consolidation message containing node identifier, current time, and adopted frequency Disseminate message

Channel Consolidation Disseminate message: For each of g available frequencies in channel list Repeat until rebroadcast overheard and at most m times Broadcast message Listen for rebroadcast Await consolidation: Listen on adopted frequency If receive consolidation message If message contains available frequency different from adopted If message contains most recent timestamp Adopt frequency Record timestamp Disseminate message

Consolidating Channels

Single Common Channel

Channel Negotiation Assumptions for each node: - initial spectrum sampling concluded - adopted frequency - list of frequencies of all common channels overheard - information communicated as packets - notion of common time - at least one node must adopt a different frequency Initiate negotiation: If adopted frequency no longer inappropriate Select least-busy available frequency different from adopted Record frequency in list of channels Adopt frequency Generate negotiation message containing node identifier, current time, adopted frequency, and list of unavailable frequencies Disseminate message

Channel Negotiation Await negotiation: Listen on adopted frequency If receive negotiation message If frequency available and different from adopted If message timestamp most recent Adopt frequency Record timestamp Update known unavailable frequencies and add to message Disseminate message If frequency unavailable If message timestamp most recent If common available frequency exists Generate negotiation message containing node identifier, current time, selected frequency, and list of unavailable frequencies Disseminate negotiation message

Negotiating New Channels

Clusters of Common Channels

Selecting Next Hops for Packet Forwarding Assumptions for each node: - channels established - neighbors discovered - hop-by-hop acknowledgements - information communicated as packets Algorithm: - objective: low costs routes at low overhead - quasi dynamic programming to compute cost to destination - probabilistic selection of packet’s next hop to destination - probabilities computed according to learning automata rules or strictly based on cost - positive probabilities prevent sticking - forwarding loops can occur but do not persist - transmission cost of loops less than that of state information in dynamic networks

Probabilistic Packet Forwarding Select next hop to destination d via neighbor j according to probabilities {pid(t): 1 ≤ i ≤ m} for all m neighbors Forward data packet to neighbor j and obtain in acknowledgement estimated cost cjd(t) to destination d Update estimate of cost to destination d via neighbor j: cnjd(t+1) = linknjd(t) + cjd(t) Update probability of selecting each neighbor k as next hop to destination d: pkd(t) = (1/ cnkd(t)) / ∑(1/ cnid(t)) over all m neighbors Update estimate of cost to destination d: cnd(t) = ∑ pid(t) cnid(t) = m / ∑(1/ cnid(t)) over all m neighbors

Variant Based on Learning Automata Update probability of selecting each neighbor k as next hop to destination d according to linear reward-penalty scheme but constrained positive: pjd(t+1) = pjd(t) + a(1 - ß(t))(1 - pjd(t)) - bß(t)pjd(t) pkd(t+1) = pkd(t) - a(1 - ß(t))pkd(t) + bß(t)(1/(m-1) - pkd(t)), k ≠ j Reward: ß(t) = 0, route via neighbor j has lowest cost Penalty: ß(t) = 1, route via neighbor j does not have lowest cost 0 < b << a << 1 Update estimate of cost to destination d: cnd(t) = ∑ pid(t) cnid(t), over all m neighbors.

Forwarding Packets

Forwarding Packets