Network Wide Broadcast in MultiHop Wireless Networks Some
Network Wide Broadcast in Multi-Hop Wireless Networks Some slides adapted from Broadcast Storm author’s presentation 1
Network Wide Broadcast Problem Network Wide Broadcast (NWB): NWB used for Try to take advantage of MAC level broadcast send a message to all nodes in the network May not have topology information Routing Paging Group communication One transmission covers multiple receivers Don’t have to know who is in range with you 2
Preliminaries The broadcast is spontaneous. The broadcast is unreliable. no synchronization no prior global topology knowledge – we relax this later no acknowledgement of any kind • Why? Think about it • not to cause more contention • 100% reliability is unnecessary for some application Performance Metrics? Overhead—does it make sense? How about normalized overhead? Robustness—how to measure that? (coverage? ) 3
Saving Rebroadcasts 5 forwarding nodes 4 hop time source Can cover all—but extra Latency and lower reliability 6 forwarding nodes 3 hop time source Flooding—high overhead, But shortest paths? 4
NWB by Flooding A straight-forward approach A host rebroadcasts the message on receiving a broadcast message for the first time. Broadcast storm problem: But maybe the redundancy is useful to improve robustness? Next time we examine this angle redundant rebroadcasts contention problem collision problem 5
Analysis of Redundancy Additional Coverage provided by a rebroadcast. The max. additional coverage is 61%. The coverage is 41% in average. The expected additional coverage EAC(k)/ r 2 after a host heard a broadcast message k times. 6
Analysis on Contention When a host broadcasts, its neighbors are likely to contend with each other for the medium. Analysis on Contention A ==> B, C, D could seriously contend with each other. A B C D 7
Broadcast Losses Higher Possibility of Collision and broadcast loss: Rebroadcasts are likely to start at the same time. • Backoff window runs out if medium is quiet for a while. hidden terminal problem Wireless propagation losses No Acknowledgements and retransmits—losses are fatal we get only one shot What are the implications? 8
Broadcast Storm Problem Summary Redundancy Contention Collision and losses How to derive an efficient scheme for broadcasting in a MANET? 9
Possible Broadcast Solutions Probabilistic Scheme Counter-Based Scheme Distance-Based Scheme Location-Based Scheme Cluster-Based Scheme 10
Serious Redundancy Optimal broadcasting vs. Flooding Severity of Redundant Coverage. (a) optimal = 2 steps (b) optimal = 2 steps 11
Probabilistic Scheme Rebroadcast by “Tossing a Die” A host always rebroadcasts with a certain probability P. When P = 1, this is flooding. A smaller P will reduce the storm effect. This approach is also known as gossiping in distributed systems It has well known and analyzed properties 12
Simulation Parameters no of hosts = 100 transmission radius = 500 meters packet size = 280 bytes transmission rate = 1 M bits/sec broadcast arrival rate: 1 per sec. to the whole map: (1 unit = 500 meters) roaming pattern: random walk 1 x 1, 3 x 3, 5 x 5, 7 x 7, 10 x 10 speed: 0~10 km/hr in a 1 x 1 map, 0~30 km/hr in a 3 x 3 map, etc. IEEE 802. 11 13
Performance of Probabilistic Scheme RE = REachability (in lines) SRB = Saved Re. Broadcast (in bars) Latency 14
Observation Reachability: In smaller maps, a low P is sufficient to achieve high reachability. A larger P is needed in a larger map. Saved Rebroadcast: Latency: linear with respect to P Interestingly, in smaller areas, broadcast tends to complete in a slower speed. 15
Counter-Based Scheme If a host has received a broadcast packet > C times, Examples: Addition Coverage then do not rebroadcast. 1 time => 41% 2 times => 19% 3 times => 9% 4 times => 5% > 4 times, very little extra area 16
Performance of Counter-Based Scheme We vary C = 2, 3, . . . , 6 to observe the performance. A larger C means more rebroadcast. 17
Observation Reachability: Saved Rebroadcast: C >= 3 can offer a reachability close to flooding. In denser area, there is more saving. In sparser area, there is less saving. Latency: Higher latency is smaller area. 18
Distance-Based Scheme Calculate the distance to the sending host. dmin = Min{the distance to each sending host} If dmin < D (a threshold), then do not rebroadcast. How to find distance: signal strength GPS devices 19
Performance of the Distance-Based Scheme We vary D = 147, 72, 37, 20, 11 to observe the effect. Smaller D means more rebroadcasting. 20
Observation Why choosing D=147? Reachability: Saved Rebradcast: Latency: addition coverage = 0. 187, equal to that of C=2 All look good, close to flooding. not much smaller area has higher latency 21
Location-Based Scheme From GPS to obtain the sender’s location. Let (x 1, y 1), (x 2, y 2), (x 3, y 3), . . . , (xk, yk) be locations of senders. We can accurately calculate the additional coverage of this rebroadcast. 22
Difficulty Involve complicated math to calculate the extra coverage. A lot of calculus! Approximation: grid simulation S 1 A S 3 S 2 23
Performance of the Location-Based Scheme We vary A (addition coverage) from 0. 1 to 0. 01. Smaller A means more rebroadcast. 24
Observation Why choosing A=0. 187? Best performance over all the above schemes! This is additional coverage offered by C=2. 25
Modified Location-Based Schemes Polygon Test: If a node is within the polygon formed by the locations of senders, then DO NOT rebroadcast. (Fig. (a)) Otherwise, rebroadcast. (Fig. (b)) If a host is within the convex, the maximum additional coverage is well below 22%. (Fig. (c)) 26
A Short Summary Main Concern: Different levels of accuracy: Probabilistic not sensitive to the importance of a retransmission Extra coverage of a rebroadcast probabilistic, counter, distance, location, polygon Maybe ok in dense networks Others are sensitive, with increasing accuracy of estimate (and increasing amount of work needed) Counter-Based Scheme < Distance-Based Scheme < Location-Based 27
Cluster-Based Scheme Cluster formation algorithm Each host has a unique ID A host with a local minimal ID will elect itself as a cluster head. This head host together with its neighbor will form a cluster. These neighbor hosts are called member of the cluster. 28
Cluster-Based Scheme Cluster formation protocol: A head’s rebroadcast can cover all other hosts in that cluster if its transmission experience no collision. Gateway hosts should take the reponsibility to propogate the broadcast msg to hosts in other clusters. There is no need for a non-gateway member to rebroadcast the msg. 29
Follow up work by the same authors (ICDCS 2001) In the above solutions, the thresholds used are ALL FIXED. NOT sensitive to the current status of the network. Example: In the counter-based scheme, we may need different threshold C depending on the density of the network. Why not make threshold adaptive to the network status? They explore this idea– it works fine 30
Brief Taxonomy of Approaches Topology Ambivalent (or flooding based) Another approach is possible – Topology aware No knowledge of neighbors assumed Starting point is flooding These are the broadcast storm solutions we’ve seen so far General approach: use network topology information to intelligently decide who should rebroadcast Cluster based approach is an example Centralized solution: figure out the minimum set of rebroadcasting nodes to cover all (Minimum Connected Dominating Set MCDS) Centralized– unrealistic Does not consider that rebroadcasts are unreliable • • • 31
Minimum Forwarding Set Problem Define: Given a source A let D and P be the sets of k and k+1 hop neighbors of A Find a minimum-size subset F of D such that every node in P is within the coverage area of at least one node from F In general graph: In unit disk graph: NP-complete: reduce “Set Cover” to it Approximation ratio: logn Unknown Approximation ratio: constant 32
Minimum Broadcasting Set Problem Define: Given a source A Find a spanning tree T such that the number of internal nodes is minimum In general graph: In unit disk graph: NP-hard: hard to “Minimum Connected Dominating Set” Approximation ratio: log ( is the maximum node degree) NP-hard Approximation ratio: constant 33
Hiearchical: Domination-set-based School bus routing Next few slides from Prof. Jie Wu at FAU 34
Graph-theoretic Definition A set in G(V, E) is dominating if all the nodes in the system are either in the set or neighbors of nodes in the set. 35
Five-Queen Problem (1850’s) 36
Desirable Features Simple and quick Connected dominating set Figure 6: A simple ad hoc wireless network of five wireless mobile hosts. 37
Existing Approaches Graph theory community: Bounds on the domination number (Haynes, Hedetniemi, and Slater, 1998). Special classes of graph for which the domination problem can be solved in polynomial time. 38
Existing Approaches (Cont’d. ) Ad hoc wireless network community: Global: MCDS (Sivakumar, Das, and Bharghavan, 1998). Quasi-global: spanning-tree-based (Wan, Alzoubi, and Frieder, 2002). Quasi-local: cluster-based (Lin and Gerla, 1999). Local: marking process (Wu and Li, 1999). 39
MCDS (Sivakumar, Das, and Bharghavan, UIUC) All nodes are initially colored white. The node with the maximum node degree is selected as the root and colored black. All the neighbors of the root are colored gray. Select a gray node that has the maximum white neighbors. The gray node is colored black and its white neighbors are marked gray. Repeat step (3) until there is no more white node. 40
MCDS (Cont’d. ) black nodes = CDS (connected dominating set) Figure 7: MCDS as an approximation of CDS 41
Spanning-tree-based (Wan, Alzoubi, and Frieder, IIT) A spanning tree rooted at v (selected through an election process) is first constructed. Nodes are labeled according to a topological sorting order of the tree. 42
Spanning-tree-based (Cont’d. ) Nodes are marked based on their positions in the order starting from root v. All nodes are white initially. V is marked black and all nodes are labeled black unless there is a black neighbor. Each black node (except root v) selects a neighbor with the largest label but smaller than its own label and mark it gray. 43
Spanning-tree-based (Cont’d. ) black nodes = DS black nodes + gray nodes = CDS Figure 8: selecting CDS in a spanning tree 44
Cluster-based (Lee and Gerla, UCLA) All nodes are initially white. When a white node finds itself having the lowest id among all its white neighbors, it becomes a cluster head and colors itself black. All its neighbors join in the cluster and change their colors to gray. 45
Cluster-based (Cont’d. ) Repeat steps (1) and (2) until there is no white node left. Special gray nodes: gray nodes that have two neighbors in different clusters. 46
Cluster-based (Cont’d. ) black nodes = DS black nodes + special gray nodes = CDS Figure 9: sequential propagation in the cluster-based approach. 47
Localized Algorithms Processors (hosts) only interact with others in a restricted vicinity. Each processor performs exceedingly simple tasks (such as maintaining and propagating information markers). Collectively these processors achieve a desired global objective. There is no sequential propagation of information. 48
Marking Process (Wu and Li, 1999) A node is marked true if it has two unconnected neighbors. A set of marked nodes (gateways nodes) V’ form a connected dominating set. 49
Marking Process (Cont’d. ) Figure 10: A sample ad hoc wireless network 50
Dominating Set Reduction Reduce the size of the dominating set. Role of gateway/non-gateway is rotated. 51
Dominating Set Reduction (Cont’d. ) N [v] = N (v) U {v} is a closed neighbor set of v Rule 1: If N [v] N [u] in G and id(v) < id(u), then unmark v. Rule 2: If N (v) N (u) U N (w) in G and id(v) = min{id(v), id(u), id(w)}, then unmark v. 52
Dominating Set Reduction (Cont’d. ) Figure 12: two sample examples 53
Example Figure 13: (a) Dominating set from the marking process (b) Dominating set after dominating set reduction 54
Topology Aware Approaches Neighborhood information How to decide forwarding nodes Dynamically, or Neighborhood base • Scalable Broadcast Algorithm (SBA), Flooding with Self pruning Statically, or Set cover base Multipoint relaying, Dominant pruning, Ad hoc Broadcast Protocol (AHBP) MCDS base • 55
Scalable Broadcast Algorithm (SBA) Information: Hello message (2 -hop) Forwarding node decision: Node vj who receives the packet from vi checks whether the set N(vj)N(vi)-{vi} is empty Node vj schedules the packet for delivery with a RAD (Random Assessment Delay) Dynamically adjust the RAD to (nodes with the most neighbors usually broadcast before the others) 56
Self pruning Information: Hello message (1 -hop) Piggyback adjacent node list in broadcast packets (2 -hop) Store adjacent node list in cache Forwarding node decision: Node vj who receives the packet from vi checks whether the set N(vj)-N(vi)-{vi} is empty vi vj 57
Multipoint relaying Information: Forwarding node decision: Hello message (2 -hop) The sending node A selects forwarding nodes from it’s adjacent nodes A select a minimum node set F N(A) such that: A node set U = N(N(A)) – N(A) Piggyback forward list in “Hello” packets 58
Dominant pruning Information: Forwarding node decision: Hello message (2 -hop) The sending node selects forwarding nodes from it’s adjacent nodes Node vj who receives the packet from vi , vj select a minimum node set F N(vj) N(vi) such that: A node set U = N(N(vj)) – N(vi) – N(vj) Piggyback forward list in broadcast packets 59
Dominant pruning N(N(vj)) N(vi) N(vj) B(vi, vj) U 60
The drawback of present set cover based protocols 1 When a node vi received the broadcast packet from node vi-1, it will select some forwarding nodes from N(vi)-N(vi-1) to cover all nodes in U. However, some nodes in U are not i+2 level nodes, and some nodes in N(vi)-N(vj) are not i+1 level nodes. vi vi-1 s … i-1 i i+1 i+2 61
The drawback of present set cover based protocols 2 When we will select some level i+1 nodes to cover all level i+2 nodes, the number of forwarding nodes selected by distributed algorithm can not be bounded to some ratio of the optimal solution ? vi 2 vi-1 s vi 1 … i-1 i i+1 i+2 62
350 x 350 Comparison r: 100 63
Robustness of Network Wide Broadcasts (NWBs) 64
NWB Algorithms--Recap Optimized NWB algorithms cut down on the number of broadcasts – Redundancy control Topology Aware vs. Topology Ambivalent Aware: Collect neighbor information, construct a virtual backbone (CDS Approaches) Ambivalent: Start with flooding, cut down on some broadcasts (based on local criteria) Tradeoffs: Aware can reduce redundancy, has additional neighbor-tracking overhead 65
NWB Algorithms - 2 Static vs. Dynamic (Topology-Aware) Static – Forwarding nodes pre-selected by previous hop Dynamic – Nodes locally decide whether to rebroadcast (calculates whether some neighbors are uncovered) Tradeoffs – Static algorithms are more optimal (lower overhead), dynamic algorithms adapt to losses better 66
NWB Robustness Problem Optimizing NWBs cuts down on redundancy Key broadcasts can be lost, isolating sections of the network Metrics of interest: Coverage – Percentage of reachable network covered Normalized Overhead – Cost associated with the NWB operation 67
Simulation Information NS-2, 802. 11 MAC protocol 20 scenarios of 30 nodes (randomly distributed) – 20 random seeds for each scenario Static scenarios (some tests done with probabilistic random walk) Propagation models Two Ray Ground Shadowing – Idealized & harsh signal fading model 68
Characterization of NWB Robustness - Flooding Coverage 69
NWB Robustness Protocols studied: Flooding Location-Based Algorithm – from broadcast storm paper • Topology ambivalent Ad Hoc Broadcast Protocol (AHBP) Static CDS Scalable Broadcast Algorithm (SBA) Dynamic CDS Double-Covered Broadcast (DCB) – Static CDS with double coverage (Infocom 2004) Only topology-aware algorithm with robustness control • • 70
NWB Robustness Trends Sparse networks are more prone to losses Topology-aware protocols are not as robust when losses occur Topology-aware protocols have lower overhead Static CDS approaches are not as robust as dynamic approaches 71
Coverage of NWB Protocols 72
NWB Coverage (Mobility) 73
Robustness Control Broadcast Storm optimize to reduce redundancy Losses figure out when to introduce redundancy to compensate Robustness control An effective NWB must include both aspects Core issue is MAC level broadcast is unreliable Robustness vs. Reliability 74
Solution Space for Robustness Control Loss-sensitive vs. fixed redundancy Loss-sensitive • implicit or explicit loss detection? Fixed redundancy: • State-sensitive or blind redundancy? Network level vs. MAC level implementation Network level easier to implement and deploy • +can take wider range of measures in space or time • -But must build on what exist at MAC level 75
Explicit Feedback Receivers directly receive feedback information regarding the NWB success Granularity of feedback: Every packet or periodically? Number of reports: Every node, some nodes, or a single node? 76
Implicit Feedback Predict losses based on observed behavior Difference between a reception and a loss is the subsequent rebroadcasts No explicit feedback Observe overheard rebroadcasts Can do better with some state information on expected rebroadcast behavior 77
MAC Primitive by Pei and Gerla (circa 1999) Not directly focused at NWB—rather, they just want to make MAC broadcast more reliable Idea, Broadcast, but require acknowledgements No ACKs? Rebroadcast One ACK? Good, at least one person received it ACK explosion? • There will be noise on the channel due to ACK collisions • Optimistically guess that it is ACKs Ensures at least one receiver Requires What is this approach in terms of the solution space? How will it work? 78
Another MAC Primitive--Directed Broadcast Problem—many potential receivers; difficult to carry out ACK/retransmit with multiple partners Directed Broadcast: pick a partner to ACK Also does not ensure that all receive Partner responds with ACK Others receiver but do not ACK “Public Unicast” But minor modifications to MAC; no new physical abilities needed Reduces load on network (not all need to ACK) Allows use of RTS/CTS if desired 79
Directed Broadcast Gives control of who is guaranteed to receive this can be exploited by topology-aware NWB algorithms • E. g. , CDS algorithms can pick a CDS member as partner • Works very well for these algorithms But, is more difficult to use by topology-ambivalent algorithms You need to know at least one partner What type of solution is it? 80
Receiver Driven Sequence Numbers (Gerla) NWB originator uses sequence numbers for its broadcasts Each receiver keeps the last few broadcasts When a receiver receives a packet from one of its neighbors, and it discovers a “hole” in the sequence numbers, it asks for a rebroadcast of the missing packet Explicit feedback, NAK based Big drawback Can it ensure full reliability? Delay bounded? 81
Variations/Other possibilities Stagger ACKs to allow multiple responders Directed Broadcast with multiple partners All of these (schemes + these variations) are MAC based, explicit feedback with packet level granularity Why? Tricky, especially when responders not known Variation of this exploited for Anycast—will discuss later Repeat the broadcast to a subset of nodes If all nodes, it becomes unicast Can pick all members of the CDS for example 82
Other Robustness Control Hyper-gossipping/Hyper-flooding Start with flooding or gossiping, and with a probability retransmit again Network-level, fixed redundancy, not- state aware Only Gossiping and flooding supported, not the more advanced redundancy control mechanisms But maybe the other schemes can be supported too? • Double-Covered Broadcast (DCB)–Construct a CDS that is doubly covered • Each node is part of the CDS or is covered by 2 CDS members • Tradeoff vs. a traditional CDS? • Network-level, fixed redundancy, not-state aware 83
Selective Additional Rebroadcast (SAR) 84
SAR Areas of concern are situations where loss is likely and the network has low redundancy Can we use implicit feedback to speculate whether a packet has been lost or not? Can we use state information to improve our guess? Network-level, implicit feedback, possibly state aware Basic idea: Implicit loss detection Retransmit if loss is suspected Can be layered on top of any “redundancy control” NWB approach 85
Loss Prediction Approaches - 1 Probabilistic A node that broadcasts will perform a second broadcast based on some probability p Does not use any feedback, resulting in poorer results Blindly adds robustness • • Counter-based A node that broadcasts waits to hear if the packet was sent by n neighbors Uses implicit feedback Not blind, but not adaptive • • 86
Loss Prediction Approaches - 2 State-Aware (Adaptive) Disable SAR if past broadcasts have not been helpful Neighbor-knowledge protocols can ensure rebroadcasting nodes heard the broadcast Observe MAC utilization 87
SAR Approaches Comparison Probabilistic – Fixed level of overhead, some gain in coverage, some rebroadcasts are wasteful, some needed rebroadcasts were missed Counter-Based – Gain in coverage, not fullyoptimized in terms of overhead (leaf nodes tend to rebroadcast) Adaptive – Results at or above counter-based, lower overhead. Adjusts to network conditions 88
SAR Results 89
SAR Overhead 90
Link Quality Sensitive CDS 91
Signal Fading Chance of packet reception is based on a number of factors Transmit power Distance between nodes Obstacles in propagation path 92
Packet Reception 93
Link Quality Aware CDS Like DCB, this is a fixed-redundancy & topologyaware approach Sensitive to link states • One good link may be better than two poor links Build CDS that keeps weights on each path Attempt to cover all nodes with some probability P 94
Example 0. 1 A 0. 6 B 0. 6 C 95
LQ CDS Coverage 96
LQ CDS Overhead 97
LQ-CDS 1. Work in progress; obviously needs a lot of work Problems: Shadowing model is too harsh Most links have extremely poor quality—network effectively very poorly connected Should consider a less harsh model or denser scenarios where there at least some ok links to work with 2. Does not consider that the probability of coverage is conditional on all the sources being covered Heuristic Solution: increase retransmit probability with number of hops from source Track the cumulative probability 98
- Slides: 98