Presentation is loading. Please wait.

Presentation is loading. Please wait.

Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly Department of Computer Science University of Pittsburgh And John Augustine.

Similar presentations


Presentation on theme: "Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly Department of Computer Science University of Pittsburgh And John Augustine."— Presentation transcript:

1 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly Department of Computer Science University of Pittsburgh And John Augustine Donald Bren School of Information and Computer Sciences University of California at Irvine

2 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 2 Agenda General network model. Previous results for general networks: Oblivious routing results. Admission control results. Our Results for sensor networks: Lower Bounds. Logarithmic oblivious routing algorithm for trees. Logarithmic oblivious routing algorithm for 3D grid networks.

3 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 3 Model of General Networks We model the network as a graph G=(V,E) of |V| = n nodes and |E| edges. Edge weights = link bandwidths. For packet-switched networks, a set of demands D ab ≥ 0 for each origin-destination (OD) pair (a, b). A routing algorithm f: determines an s-t path for each unit demand of each OD pair: f ab ( i,j ) ≥ 0 routing for OD pair (a,b) on edge (i,j) Flow on edge e=(i,j) when routing D with f: flow(e,f,D)=  ab f ab ( i,j ) D ab

4 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 4 Performance Metrics for Routing Algorithms Congestion on edge e=(i,j) when routing D with f: cong(e,f,D)=flow(e,f,d)/capacity(e) Congestion of demands D with routing f: cong(f,D)= max cong(e,f,D), for all e  E Optimal Offline Algorithm OPT: has full knowledge of the input set D and can process it optimally. Competitive Worse-case Analysis: compare the congestion cong(f,D) of any algorithm f with that of the optimal offline algorithm OPT cong(OPT,D) and bound it for all possible inputs D: cong(f,D) ≤  c  cong(OPT,D) + a  for any D c is called the competitive ratio of algorithm f. High probability argument for randomized algorithms

5 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 5 Routing Algorithms under Concern Online routing algorithms: demands D arrive in an online fashion. For each request, the algorithm determines the path without knowing future requests. Oblivious routing algorithms: The path chosen for a unit demand is independent of the current network load. In other words, the routing path of a message (s,t) only depends on s,t and does not depend on all other messages. We are interested in routing algorithms that are: Online. Distributed. Oblivious Routing. Poly-logarithmic competitive ratio (w.r.t congestion).

6 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 6 Related Problem: Online Call Control Problem definition: A sequence of requests arrive in an online manner. For each request, the algorithm A has to either accept or reject the call. When accepting a call, the algorithm has to immediately select a virtual circuit between the communicating parties, obeying the network constraints. Arises in the context of circuit-switched networks. Performance Metric: number of accepted calls, i.e. the throughput of the call control algorithm, P(A,D). Similar bandwidth requirements for all calls. Competitive analysis: P(A,D) ≤  c  P(OPT,D) + a  for any D

7 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 7 Admission Control Algorithms under Concern We are interested in routing algorithms that are: Online. Distributed. Oblivious path selection step. Poly-logarithmic competitive ratio (w.r.t throughput).

8 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 8 Oblivious Admission Control & Routing Online algorithm. Distributed. Call is obliviously admitted/rejected. Path is obliviously determined. Goal of such algorithms: Maintain poly-logarithmic competitiveness w.r.t either congestion in packet networks or throughput in circuit-switched networks.

9 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 9 Oblivious Routing in General Networks Räcke, FOCS 2002: Any undirected network has an online distributed oblivious routing with competitive ratio O(log 3 n), w.r.t congestion!! Highlights of the technique: Map the network G=(V,E) into a tree T G =(V t, E t ), the so-called decomposition tree. Show that: Network  decomposition tree. Route the messages on G=(V,E) based on the usage of T G =(V t, E t ) to determine message paths.

10 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 10 More on Räcke’s Algorithm Strengths: It is the first routing algorithm that is Oblivious. Decentralized. Online. Weaknesses: Decomposition tree construction is not in polynomial time. Not applicable for directed graphs.

11 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 11 Follow-ups of Räcke’s Algorithm Azar et al., STOC 2003: A polynomial time construction algorithm for Räcke’s decomposition tree (based on the usage of the LP ellipsoid method) Bienkowski et al., SPAA 2003: A practical algorithm for the decomposition tree construction with a competitive ratio of O(log^3 n). Harrelson et al., SPAA 2003: drop approximation ratio to O(log^2 n log log n). Hajiaghayi et al., STOC 2005: oblivious routing for directed graphs which is O(log 2 n)-competitive assuming demands chosen at random from a known demand distribution. Hajiaghayi et al., SODA 2006: lower bounds on oblivious routing in undirected graphs.

12 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 12 Admission Control in General Networks Awerbuch, Azar, and Plotkin, FOCS 1993 ([AAP] algorithm): First online admission control algorithm. O(log n)-competitive. Admission and routing decisions are based on the knowledge of all network links. A central node in charge of taking decisions. “Hot spot” in networks with high traffic volumes. “Single point of failure”: operation of the whole network depend on the availability of the central node. Once a call is admitted, link identities of the links along its chosen path have to be communicated to the nodes along the path.

13 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 13 Admission Control in General Networks Räcke and Rosén, SODA 2005 ([RR] algorithm): Distributed randomized online call control algorithm. Poly-logarithmically competitive (w.r.t throughput). Handles concurrent requests in the network. Two components: Path establishment and call admission. Operates in a hop-by-hop manner. Given a request, a sequence of messages is sent along some path that is defined on a hop-by-hop basis. Path establishment based on Räcke’s decomposition tree. If a path is established, call routed on that path. Otherwise, the call is dropped. The network incurs a cost only when path is admitted.

14 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 14 Open Question Is it possible to get a poly-logarithmic oblivious routing algorithm for sensor networks (or in general, in node- capacitated undirected graphs)?? Previous results: Assuming random demands, a lower bound of Ω(log n/log log n) on oblivious routing schemes for a family of large graphs (grids) was shown by Hajiaghayi et al., SODA 2006. Is it possible to use Räcke’s decomposition or the RR algorithm to achieve an upper bound?? No, unfortunately.

15 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 15 Why Not RR Algorithm for Sensor Networks?? Sensor network: Energy-capacitated sensor nodes. Main goal of a routing algorithm: maximize number of successfully sent messages, i.e. throughput. Sending any message consumes an amount of the node’s total energy!!! Node deaths, Network partitioning, network lifetime (longevity). RR’s Path establishment messages are not free of charge as in general networks. RR Algorithm is not poly-logarithmic (w.r.t. throughput) for sensor networks. Is there a poly-logarithmic (w.r.t. throughput) admission control and oblivious routing algorithm??

16 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 16 Our Results Blend call control with oblivious routing. Lower Bounds: Any distributed deterministic packet-admission and oblivious routing algorithm cannot be poly- logarithmically competitive against adversarial demands. An always-send distributed oblivious routing algorithm cannot be poly-logarithmically competitive when: 1. Demands are either adversarial or following a general distribution that is unknown to all sensor nodes. 2. An adversary sets the tree node capacities.

17 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 17 Our Results Upper Bounds: A O(α log n)-competitive algorithm (w.r.t. throughput) for tree networks assuming adversarial demands and node capacities  [k, αk], where k is Ω(log n). A O(log n)-competitive algorithm (w.r.t. throughput) for grid networks assuming random demands as well as node capacities are k ≥ log 2 n + 6 log n.

18 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 18 Lower Bounds: Adversarial Demands Lemma 1: There exists at least one set of adversarial demands D that makes any deterministic distributed routing algorithm A at least Ω(n) competitive. Proof on balanced binary tree T. Assumptions: Adversarial demands. Adversary knows all tree node capacities at any time. Tree T follows the nesting property. Each leaf has a capacity e, such that e << n.

19 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 19 Lower Bounds: Adversarial Demands Tree T=(V,E), |V| = n. Messages from left subtree to right subtree. Bit vector of size n at each of the left subtree leaves. Message to r k is sent if bit k is set to 1. Message (s,t)  Tree  One path Adversary inputs messages till depleting all tree nodes. log n l1l1 l2l2 l3l3 l4l4 r4r4 r1r1 r2r2 r3r3

20 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 20 Lower Bounds: Adversarial Demands Main idea: make any deterministic algorithm send messages that OPT would have not sent. Adversarial strategy operates in n rounds. At each round, adversary selects at random one still alive node r k. For each left leaf l i, input messages (l i,r k ) till l i changes bit entry k to 1. log n l1l1 l2l2 l3l3 l4l4 r4r4 r1r1 r2r2 r3r3

21 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 21 Lower Bounds: Adversarial Demands For each round, first e messages deplete node r k. Rest of messages (at least n) are falsely sent by any deterministic algorithm. After e rounds, both algorithms have sent at most e 2 messages. After e rounds, online algorithm depletes all tree nodes (nesting prop.). After n rounds, OPT sends n.e messages. Thus, approximation ratio is n/e. log n l1l1 l2l2 l3l3 l4l4 r4r4 r1r1 r2r2 r3r3

22 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 22 Lower Bounds: Demands Drawn from Unknown Distribution Lemma 2: Assuming demands D are drawn from a distribution that is oblivious to all sensors, there exists a distribution that makes an always-send algorithm A as at least Ω(n) competitive. Proof on balanced binary tree T. Assumptions: n is a power of 2. Adversary knows all tree node capacities at any time. Tree T follows the nesting property. Leaf have equal capacity belonging to log n.

23 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 23 Lower Bounds: Adversarial Leaf Node Capacities Lemma 3: Assuming demands D are drawn from a uniform distribution, there exists at least one adversarial setting for leaf node capacities distribution that makes an always-send algorithm A as at least Ω(n) competitive. Proof on balanced binary tree T. Assumptions: n is a power of 2. Adversary knows all tree node capacities at any time. Tree T follows the nesting property. Lemma implies a lower bound on adversarial setting of all tree nodes rather than just leaves.

24 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 24 Main Lower Bound Theorem Theorem: Given a balanced binary tree T(V,E) and a set of demands D, an always-send oblivious routing algorithm A as cannot maintain poly-logarithmic competitiveness in either of the following cases: D is a set of adversarial demands, or follows a general distribution that is unknown to all sensor nodes. An adversary sets the tree node capacities (internal nodes or leaf nodes). Results extends to general trees and general undirected graphs. Packet-admission protocol is mandatory poly- logarithmic competitiveness.

25 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 25 Upper Bound for Sensor Trees Tree T=(V,E), |V| = n Balanced with height O(log n) Each v  V has energy capacity (k+1) · log n Each message is a pair (s,t) It originates at s Its intended destination is t T is a tree  Unique path for each message

26 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 26 Routing in Trees s t Tree T=(V,E), |V| = n Message (s,t)  Tree  One path Capacity of each node is at least (k+1) log n log n k (log n) for messages (log n) for control info.

27 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 27 Routing in trees Input instance I A tree T A sequence of messages released online Message i will not be released until (i-1) is A. Rejected, or B. Accepted and routed OPT( I ) = # of messages routed by an opt. algorithm Without violating capacity constraints Let A be an algorithm that admits A ( I ) messages A is c-competitive if OPT( I )  c · A ( I )

28 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 28 Pivot Vertices p s t Every message (s,t) pivots at some p Pivot of a message = Least Common Ancestor (LCA) of its source s and destination t.

29 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 29 Our Algorithm A Each message is classified based on its LCA. Each node uses at most k units of energy (1 unit = send 1 message). Each node partitions its energy into log n shares with energy k/log n each. A node admits messages with LCA at level i as long as its i th share is not totally consumed. When a message is admitted at any node, the node decrements its share i by 1.

30 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 30 Our Algorithm A Each vertex v is pivot for up to k messages After that quota is met Broadcast a message to all descendants Inform them to reject messages pivoting at v Theorem: A is (log n)-competitive Vertex capacity Height of Tree

31 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 31 Consider one pivot v For some input instance I Let OPT(v) = # of messages OPT pivots at v A (v) = # of messages OPT pivots at v Recall: A routes first k messages OPT can route up to k log n messages

32 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 32 Consider Entire Tree Recall:

33 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 33 Capacity Constraint When a pivot vertex reaches the limit Broadcasts to descendants: reject further messages. Each vertex can receive at most log n such control messages from its ancestors Each vertex can participate in ≤ k(log n) messages # pivoting at one ancestor # of ancestors (at most)

34 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 34 Properties of A Distributed: No central node for call control Oblivious Decisions can be made without detailed knowledge of the rest of the tree. We account for every message hop

35 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 35 Extending Algorithm for General Trees  a partition vertex such that # of vertices in the bigger partition ≤(2/3)n Recursively partition the tree and generate subtrees. Each tree gets an equal quota. Message (s,t) uses quota from smallest subtree it fits. Shown to be O(log n)-competitive.

36 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 36 Call Control in Grids Assumption: message generated between nodes chosen uniformly at random. Techniques: Quadtree style partitioning of the mesh. Partition each dimension into levels. Assign each level an energy quota. Route message if quotas of its 3 levels are still non- consumed. Balls in Bins (take advantage of random I/P) Result: O(log n)-competitive distributed oblivious algorithm for 3D grids.

37 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 37 Conclusion & Future Work Oblivious routing and admission control in general networks. Our results: Lower Bounds. Logarithmic oblivious routing algorithm for trees. Logarithmic oblivious routing algorithm for 3D grid networks. Future work: Upper bound for oblivious routing on general graphs. Use our theoretical results to present a new robust routing algorithm for large-scale sensor networks.

38 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 38 Acknowledgment Our advisors: Kirk Pruhs and Sandy Irani. Harald Räcke (TTI) and MohammadTaghi Hajiaghayi (CMU). NSF grants: ANI-0325353 CCF-0448196 CCF-0514058 IIS-0534531 CCF-0514082

39 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 39 Thank You Questions ?

40 Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly and John Augustine, ISAAC 2006 40 More on Räcke’s Decomposition Tree c f a b d e h i g j f bat 3 11 g hj 3 4 2 3 e 1 2 5 3 3 4 dc 1.Cluster nodes (based on bulding a laminar system). 2.Assigning capacities to tree edges. 3.Define bandwidth ratio and weight ratio for each tree level. 4.Initialize network by solving a CMCF problem. 5.For every demand, randomly select one of the paths resulting from the CMCF solution.


Download ppt "Online Packet Admission and Oblivious Routing in Sensor Networks Mohamed Aly Department of Computer Science University of Pittsburgh And John Augustine."

Similar presentations


Ads by Google