1 / 47

A Self-Coordinating Approach to Distributed Fair Queueing in Ad Hoc Wireless Networks

A Self-Coordinating Approach to Distributed Fair Queueing in Ad Hoc Wireless Networks. Haiyun Luo, Paul Medvedev, Jerry Cheng, and Songwu Lu UCLA Computer Science Department Infocom 2001. Outline. Abstract INTRODUCTION MODELS AND ISSUES A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING

leiko
Download Presentation

A Self-Coordinating Approach to Distributed Fair Queueing in Ad Hoc Wireless Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. A Self-Coordinating Approach to Distributed Fair Queueing in Ad Hoc Wireless Networks Haiyun Luo, Paul Medvedev, Jerry Cheng, and Songwu Lu UCLA Computer Science Department Infocom 2001

  2. Outline Abstract • INTRODUCTION • MODELS AND ISSUES • A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING • A DISTRIBUTED IMPLEMENTATION WITHIN THE CSMA/CA FRAMEWORK • SIMULATION EVALUATION • CONCLUSION

  3. Abstract • Distributedfair queueing in shared-medium ad hoc wireless networks is non-trivial because of the unique design challenges in such networks, such as location-dependent contention, distributed nature of ad hoc fair queueing, channel spatial reuse, and scalability in the presence of mobility. • This paper proposes a new distributed, localized, scalable, and efficient solution to these problems. • The effectiveness of this model is demonstrated through both simulations and analysis.

  4. I. INTRODUCTION (1) • The Trend • The convergence of wireless communication and networking have jump-started several wireless networking technologies such as MANET, blue-tooth, and sensor networks. • The Result • The emerging wireless technologies are envisioned to support a rich set of data applications, e.g., error-sensitive and delay sensitive applications, over bandwidth-constrained wireless medium. • The Goal • Providing fair and bounded delaychannel access among multiple contending hosts over a scarce and shared wireless channel has come to force.

  5. I. INTRODUCTION (2) • The Solution • Fair queueing has been a popular paradigm to achieve the goal in both wireline and packet cellular networking environments. • The Need • A fully distributed, scalable, and efficient fair queueing algorithm in the shared-channel ad hoc wireless nework. • The Issues • Location-dependent contention • Distributed nature of ad hoc fair queueing, • Channel spatial reuse • How to manage a large number of flows in a dense and mobile network graph?

  6. I. INTRODUCTION (3) • The Previous Works • In [10] & [12], each proposed an ideal centralized fair queueing modeland then designed a distributedimplementationto approximate the ideal model. • A Different Approach • To analyze an ideal centralized fair queueing algorithm • To extract the desired global properties from it • To devisedistributed and localized*solutions to achieve the desired global behavior. *Local schedulers use localflow information and localcomputation only.

  7. II. MODELS AND ISSUES (1) • A. Network Model • Two flows are contending with each other if either the sender or the receiver of one flowis within the transmission range of the sender or the receiver of the other flows. • Three assumptions • A collision occurs when a receiver is within the reception range of two simultaneously transmitting nodes. • A node cannottransmit and receive packets simultaneously. • Neighborhood is cumulative; thus flow contention is also communicative. • We do not consider non-collision-relatedchannel errors. • We only consider fixed packet size. But the proposed model works equally well for variable-packet-size case.

  8. ˇ ˇ x x x x x x ˇ ˇ ˇF1 and F4 are not contending! II. MODELS AND ISSUES (2) • B. Issues of Fair Queueing in Ad Hoc Networks • (a) Location-dependent contention • Since wireless communications are locally broadcast,collisions and contentions are location dependent. ∵within tx range of sender or receiver ∴contention

  9. II. MODELS AND ISSUES (3) • channel spatial reuse • ∵location-dependent contention & multihop nature • ∴Any two flows that are not interfering with each other can potentially transmit data packets over the physical channel simultaneously. • E.g. F1 & F4 • New issues to packet scheduling in ad hoc networks • In a wire-line network: • Packets are scheduled independently at each output link. • Locality property: Consider only flows that are contendingfor that link,no coordination is needed with neighboring nodes. • Fair queueing algorithms ensurelocal fair sharing of bandwidth in the time domain. • In a shared-mediummultihopwireless network: • Fair queuing becomes a 3-dimensional problem:in both the time and spatial domains. • It needs coordinated design among neighbors.

  10. II. MODELS AND ISSUES (4) • (b) Distributed nature of ad hoc fair queueing • In a wire-line network: • A switch has direct access to to the exact flow information, e.g., which flow has outstanding packets • In a cellular packet network: • The base station is the natural choice for the scheduler entity • In a shared-mediummultihopwireless network: • No single logical scheduler is available. • The flow information is spread out in the sending nodes. • The sender does not have direct access to other flows’ information at other senders.

  11. II. MODELS AND ISSUES (5) • C. Design Requirements for Fair Queueing in Ad Hoc Networks • The solution must be fully distributed, and it involves only local computations by using local information only. • The solution must exhibit desired global behavior, e.g., fairness property. • The solution must be scalable. • The solution must be efficient. • The design must be coordinated among interacting nodes.

  12. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (1) • A. Desired Global Properties • A.1 Flow graph versus node graph • To convert packet flows in a generic network topology in to a flow graph • To characterize the spatial-domain, as well as the time-domain, contention relationship among transmitting flows • Vertex: a backlogged flow • An edge between two vertex:contending flows • No edge in between:two flows can transmit simultaneously(spatial reuse)

  13. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (2) • A.2 What a centralized fair queuing model can do • Start-time Fair Queueing (SFQ) • Each arriving packet is assigned two tags: • A start tag • A finish tag • A packet with sequence number k of flow farriving at time A(tkf) is assigned two tags: a start tag Skf and a finish tag Fkf, defined as follows :where Lp denotes the packet size in bits, rf is the flow weight, and V(.) is the system virtual time, taken to be the start tag of the packet currently being serve in the scheduler during any busy period. • SFQ selects the flow with the minimum service tag (i.e. the start tag) at the moment and transmits its head-of-line packet.

  14. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (3) • Issues for adapting SFQ to the ad hoc networking environment • On tagging: • The same tagging mechanism still can be used. • However, the propagation of system virtual timeV(t) to every flow in the network is a non-trivial issue. • Scheduling: • (a) To select the flow fmin with the smallest start tag for the next packet transmission: • [10] selects flows that are not interfering with fmin, starting from those with the smallest service tags at t. • (b) To select multiple non-interfering flows for concurrent transmission: • [12] seeks to maximizespatial reuse and selects flows that solve a corresponding minimum-coloring problem.

  15. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (4) • A.3 Existing approach to approximating the centralized algorithm • Based on the framework of CSMA/CA MAC protocol [10][12] • On tagging: • If each flow can “hear” or learn an approximation of V(t) in the system, then (Eq. 1) can be used. • × V(t) is not global, i.e., at most two-hop (sender and receiver) neighborhood. • Scheduling: • Set a backoff intervalBf (t) = η* (Skf – V(t)),where Skf is the service tag of the head-of-line packet of flow f and ηis a constant. • Select the flow with the smallest Bfand transmit after Bf. • Contending nodes retains from sending through carrier sense. • Non-contending nodes can transmit concurrently.

  16. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (5) • Four limitations • (a) Local virtual time V’ (t) versus global system virtual time V(t): • Each node can only learn at best the start tag of the transmitting packet in its neighborhood. • (b) Problems with backoff setting: • If Bf is significantly large or very small, either large system overhead or potential collision occurs. • (c) Global synchronization: • All flows have to tick their backoff timer simultaneously. • (d) Potential unbounded flow unfairness in the system: • Certain flows will always transmit more often than others due to channel spatial reuse

  17. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (6) • A.4 Extracting desired global properties • Minimum fair share • Each flow is served in proportional to its flow weightrf in the local scheduler. • The flow with the globalminimum service tag at t should always transmit before other flows. (i.e. “minimize global minimum”) • Increase channel spatial reuse • Encourage concurrenttransmissions as much as possible. • Bounding unfairness • The aggregate services of two flowsf and g must satisfy where δ is a constant • Scaling property • Maintain only localinformation and perform only localscheduling computation

  18. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (7) • B. A Self-Coordinating Fair Queueing Model • B.1 Overview of the proposed model • Each node maintains a local table.The table records information of all flows in its one-hop neighborhood in the flow graph. • SFQ is used to assign start tags and finish tags.

  19. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (8) • Three mechanisms to achieve the global properties through coordination of local interactions at each node. • (1) Maximizing local minimum by transmitting flows with local minimumservice tags: • a node immediately transmits a flow only if this flow has the minimum service tag flow_tag among all backlogged flows in the table. • (2) Using the backoff mechanism to increase spatial reuse: • If a flow does not have the local minimum service tags, the backoff timer is still used. Once the timer expires and the channel is idle, the flow will be transmitted. • (3) Using a sliding window to limit the unfairness bound as an option: • Each node may maintain an upper bound δ. Whenever any flow’s service tag reaches beyond δ, the flow is retained from transmssion.

  20. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (9) • B.2 Three mechanisms in the model • (1) Maximizing local minimum • The flow with the globalminimumservice tag in the entire network must be selected for transmission before any other flows. • But, how to use local information and local computation to achieve the global information. • Since the global minimamust be a local minima (but not vice versa), the local minima must be guaranteed to be transmittedfirst. • A 3-D model • 2-D: spatial domain • 1-D: time domain • Local minimum (△) • A rain storm floods the 3-Dterrain. • Once a lowest spot is filled,it starts to fill other lowest spot.

  21. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (10) • (2) New backoff mechanism to increase spatial reuse • The “maximizing local minimum” policy alone may result in very low aggregate network utilization in a large network topology. • For any flow f, set its backoff values to be the total number of flows that have smallerservice tags (i.e., giving higher priority) than flow f. • (3) Sliding windows to limit flow unfairness if needed • Due to spatial reuse, F5 can always transmit simultaneously with other flows. Thus F5 may become unbounded. • When flow f’s service tag exceedsthe windows, it stopsin order to bound unfairness.

  22. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (1) • C. Model Description • MLM-FQ (Maximize-Local-Minimum Fair Queueing) • Implements “maximize-local-minimum” policy • EMLM-FQ (Enhanced MLM-FQ) • + “backoff-based spatial reuse” mechanism • BFMLM-FQ (Bounded-Fair MLM-FQ) ▼ • + “bounded fairness” mechanism

  23. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (2) • BFMLM-FQ • (1) Local state maintenance • Each node n maintains a local tableEn, which records each flow’s current service tag for all flows in its one-hop neighborhood. • Each entry entry has the form of [ f , Tf ], where Tf is the current service (start) tag of flow f. • (2) Tagging operations • Tag assignment: • For the head-of-line packet k of flow f, whose arrival time is A(tkf), packet size is Lp, start tag Skf, and finish tag Fkf. • (a) If f is continuously backlogged, then Skf = Fk-1f; Fkf = Skf + Lp/ rf • (b) If f is newly backlogged, then Skf = max{Vg(A(tkf)}; Fkf = Skf + Lp/ rfwhere Vg(t) is flow g’s virtual time at t.

  24. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (3) • (3) Scheduling loop • (a) If one flow, f, which has the smallestservice tag in the table, transmit the head-of-line packet of flow f immediately. • (b) Otherwise, for each flow, if Tf < Vmin + δ, where δ is the sliding window size, Vmin is the virtual time at node n, set the backoff timer Bf = ΣI(Tg(t) < Tf(t)) minislots, where I denotes the indicator function, i.e., I(x) = 1, if x > 0; I(x) = 0, otherwise. • (c) If Bfexpires, and the channel is idle, transmit the head-of-linepacket of flow f. (4) Table updates • Whenever node n hears a new service tag Tg’ for the flow g, update the table entry for flow g to [g, Tg’]

  25. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (4) • Example • 4 flows: F1 ~ F4 • Dotted lines: within tx range • Initially, • V = 0 • T1 = 1 • T2 = 2 • T3 = 3 • T4 = 4

  26. III. A SELF-COORDINATING APPROACH TO AD HOC FAIR QUEUEING (5) • Tables & backoff values for each node

  27. IV. A DISTRIBUTED IMPLEMENTATION WITHIN THE CSMA/CA FRAMEWORK (1) • Practical Issues • Exchange of the table information at a flow’s sender and it’s receiver • Propagation of each transmitting flow’s updated service tag • Hidden terminal problem

  28. IV. A DISTRIBUTED IMPLEMENTATION WITHIN THE CSMA/CA FRAMEWORK (2) • A. Protocol Overview • A.1 Basic Message Exchange Sequence • Backoff-RTS-CTS-DS-DATA-ACK • Set a backoff timer for flow f to be the number of flows with tags to be smaller than the tag of flow f. (e.g. set to 0 for the smallest) • If the backoff timer of fexpireswithout hearing any ongoing transmission, it startsRTS (carrying BfR, backoff value at the receiver side). • If it overhears some ongoing transmission, it cancels the backoff timer and defers until the ongoing transmission completes. It also updates the table for the tag of the ongoing transmission. • When a node hears a RTS, it defers for one CTS transmission time.If it is the designated receiver, it checks whether BfR is greater thanorequal to the backoff value of f, it responds withCTS. • When the sender receives the CTS, it cancels all backoff times and transmits DS. • Whenever a node hears a CTS or DS, it defers until the DATA-ACK transmission completes.

  29. IV. A DISTRIBUTED IMPLEMENTATION WITHIN THE CSMA/CA FRAMEWORK (3) • A.2 Maintaining table information at both the sender and the receiver • Bf = BfS + BfRwhere BfS and BfR are backoff values according to the sender’s table and the receiver’s table respectively. • For simplicity, assume that the sender’s table and the receiver’s table do not have identical entries for the same flows. (If an identical flow indeed exists, the receiver deletes that entry.) • However, the senderdoes not have the information at the receiver table.(sol) the sender needs to estimateBfR.

  30. IV. A DISTRIBUTED IMPLEMENTATION WITHIN THE CSMA/CA FRAMEWORK (4) • The sender does not have the information at the receiver table, thus the sender needs to estimateBfR. • The ACK carries two parameters: Mg and bg, in order for the sender to estimates BfR. • Mg: how much services (in bytes) toward other flows have to be served before flow g transmits its packet in the receiver’s table. • Mg = ΣjεB(Tj – Tg)wg, where wg is flow g’s weight,Tj is flow j’s current tag,B denotes all flows that have smaller tag Tj than Tg • bg: backoff value for flow g • The sender N estimates BgR ≒ bg˙(Mg- C ˙( t -tg )) / Mgwhere C is the channel capacity.

  31. IV. A DISTRIBUTED IMPLEMENTATION WITHIN THE CSMA/CA FRAMEWORK (5) • A.3 Propagating a flow’s updated service tag • To inform neighboring nodes of new updated service tag↓Attach the tagTffor flow f in all four packets RTS, CTS, DS, and ACK. • However, the update tags in RTS and CTS are not used because RTS and CTS do not ensure a successful transmission. • Whenever collision happens, the standard random backoff algorithm is invoked.

  32. IV. A DISTRIBUTED IMPLEMENTATION WITHIN THE CSMA/CA FRAMEWORK (6) • B. Simple Overhead Analysis • Protocol efficiency • 802.11 MAC spec: • RTS = 24 bytes • CTS = DC = 18 bytes • ACK = 20 bytes • 6 backoff slots (5 bytes each) • DATA = 512 bytes • η = 83 % (i.e., 17 % overhead)which is comparable with 802.11 • Since this paper implements enhanced scheduling, chance of collisions is reduced, thus it may compensate this overhead.

  33. V. SIMULATION EVALUATION (1) • Compare • MLM-FQ • EMLM-FQ • TWO-TIER fair scheduling • ADP-COL (adaptive-coloring-based) fair queueing • FIFO with IEEE 802.11 MAC protocol • The radio model • Simulator: ns-2 • Radio range: 250 meters • Channel capacity: 2M bps • Constant Bit Rate • Data packet size: 512 bytes • Simulation period: 1000 seconds • Identical flow weights (in order to compare with 802.11)

  34. V. SIMULATION EVALUATION (2) • A. Simulation Scenario 1 (1) x    x      Throughput complex Fairness

  35. V. SIMULATION EVALUATION (3) • A. Simulation Scenario 1 (2)

  36. V. SIMULATION EVALUATION (4) • B. Simulation Scenario 2 (1)

  37. V. SIMULATION EVALUATION (5) • B. Simulation Scenario 2 (2) x    x      complex Throughput Fairness

  38. V. SIMULATION EVALUATION (6) • B. Simulation Scenario 2 (3)

  39. V. SIMULATION EVALUATION (7) • C. Simulation Scenario 3 (1)

  40. V. SIMULATION EVALUATION (8) • C. Simulation Scenario 3 (2) x    x      Throughput complex Fairness

  41. V. SIMULATION EVALUATION (9) • C. Simulation Scenario 3 (3)

  42. V. SIMULATION EVALUATION (10) • D. Simulation Scenario 4 (1)

  43. V. SIMULATION EVALUATION (11) • D. Simulation Scenario 4 (2) x x         Throughput complex Fairness

  44. V. SIMULATION EVALUATION (12) • D. Simulation Scenario 4 (3)

  45. VI. DISCUSSIONS AND RELATED WORK (1) • A. Further Comments • A.1 Node mobility • For highly mobile wireless network, any model that requires global topology information or global computation does not work. • This design: requires only one-hop neighborhood flow information and simple local computation. • A.2 Scalability • This design: • Communicating with one-hop neighbors • Minimal information

  46. VI. DISCUSSIONS AND RELATED WORK (2) • A.3 Other issues • This design works for fixed packet size, and variable packet size as well because tagging update has counted the packet size. • A.4 Comparisons between fully distributed algorithms and our localized coordination approach • This localized coordination design is a class of fully distributed algorithms. • A common distributed algorithm seeks to approximate the centralized algorithm. • This design extractsglobal properties from the centralized algorithm, then developslocal models to capture these global properties, and finally implements the local models.

  47. VII. CONCLUSION • This paper proposes a suite of new localized and fully distributedfair queueing model for ad hoc wireless networks. • This model relies on local information and local computations only. • Multiple localized schedulers coordinate their interactions and collectively achieve desired global properties such as scalability, efficiency, fairness, and increased spatial reuse. • The effectiveness of this model is demonstrated through both simulations and analysis. • Ongoing works: • To improve the design of the distributed implementations • To implement more extensive simulations • To refine the analytical bounds of the proposed algorithm.

More Related