A New Approach to Channel Access Scheduling in Ad Hoc Networks Lichun Bao School of ICS University of California, Irvine J.J. Garcia-Luna-Aceves School of Engineering University of California, Santa Cruz Existing Solutions for Channel Access: • Random Access Scheme: – ALOHA, CSMA/CA (FAMA, MACA, MACAW, IEEE 802.11) : with/without RTS/CTS handshakes. – Difficulties providing fairness, QoS. • Scheduled Access Scheme: – Node/Link Activation. – FDMA/TDMA/CDMA in multihop networks: graph coloring problem — UxDMA. University of California 2 Our Solution: Scheduled Access • Problem description: – Given a set of contenders Mi of an entity i in contention context t, how does i determine whether itself is the winner during t ? • Topology dependence: – Exactly two-hop neighbor information required to resolve contentions. – Two-hop neighbors are acquired by each node broadcasting its one-hop neighbor set. University of California 3 Example Settings: • • • • • Omni-directional Antenna; Time slotted channel access; Equal transmission range; 4 nodes; Each node knows its one- and two-hop neighbors — Mi. University of California 4 Goals to Achieve: • Collision-free — avoid hidden terminal problem, no waste on transmissions; • Fair — the probability of accessing the channel is proportional to contention; • Live — capable of yielding at least one transmission each time slot. University of California 5 Illustrations by Example: i j l k 0 i wins 1 j & l win 2 k wins 3 time t University of California 6 Neighbor-aware Contention Resolution (NCR): • In each contention context (time slot t): – Compute priorities pkt Rand (k t ) k , k M i {i} – i is the winner for channel access if: j M i , p p t i University of California t j • 7 Attributes of NCR: • • • • Collision freedom; Fairness; Liveliness; 2-coloring: – An entity colors itself if it red has the highest priority among its contenders. – Otherwise, it has transparent color. University of California 8 NCR-MI (Multiple Identities): • Dynamic Resource Allocation. • A node i may have Ii pseudo identities. – k-th identity is denoted as i k . – Ii is dynamically chosen by i according to traffic requirement. – Each identity of i gives i a chance to win a contention. The more identities, the better chance of channel access. University of California 9 NCR-MI Specification: • Compute the priority on each pseudo identity of every contender: – For l-th identity of contender k, we have: t pk l Rand(k l t ) k l k M i {i},1 l I k • i is the winner for channel access one of its priority is the greatest among its contenders. • University of California 10 Channel Access Probability: • Dependent on the number of pseudo identities and the density of the neighborhood. • Channel access probability: – Bandwidth allocation qi Ii kM {i} I k i University of California 11 Bandwidth Allocation Example: • Channel access probability for individual 1 nodes: 1 3 j 4 i 1 4 1 3 l k • Spatial channel reuse ratio: 1 1 1 1 1.67 3 4 4 3 University of California 12 Delay & Throughput Analysis: • Data packet service at entity i modeled as M/G/1 queuing system with server vacation. • Delay (Pollaczek-Kinchin formula): i qi 2(1 qi ) 3 Ti 2(qi i ) 2 • Throughput: S min( k , qk ) k University of California 13 Delay Curves: University of California 14 Channel Access Scheduling Protocols: • Node Activation Multiple Access (NAMA): – Entity type: node – Time division: • • • • Block Section Part Time-slot University of California 15 NAMA Time Division Illustrated: Block 1 0 Section 0 ....... 50 51 Membership Section: Neighbor Maintenance 1 Part 0 Time Slot 1 2 University of California 16 NAMA Illustrated: Fully connected network with 10 nodes. ID: 1~10. Part 0 Part 1 1,5,6,8,10 2,3,4,7,9 1,10 5,8 6 4,9 2,3,7 2,3,4,7,9 Section 0 No occupied by anyone Everyone tries to use 8 1,5,6 10 9 3 2,4,7 Section 1 Contenders resolve contention using NCR University of California 17 Neighbor Protocol: • One-hop neighbor information broadcasting. – New node starting up. – Link addition and deletion. – Old neighbor going down can be treated as multiple link deletions. • Membership section: send signals. University of California 18 Channel Access Scheduling Protocols (continued): • Link Activation Multiple Access (LAMA): – Direct Sequence Spread Spectrum, available pseudonoise code set: Cpn – Received-Oriented Code Assignment (ROCA) ci c , k Rand (i) mod C pn k – Contenders of node i : M i N ( N ) {i} 1 i 1 k kn1i ,c – Once Mi is decided, LAMA follows NCR. University of California 19 LAMA Illustrated: d 1 5 14 a 11 j c 23 b e 19 Node i tries to activate its adjacent links on code c i 21 c 8 c k f 20 g i can activate either link (i,j) or (i,k). 6 University of California Both j and k are assigned code c At time t, the priority of each node is computed. 20 Channel Access Scheduling Protocols (continued): • Pair-wise Link Activation Multiple Access (PAMA): – Contending entities are directed edges; – Priorities are computed for each link; – Dynamic code assignment: cu c , k Rand (u t ) u mod C pn k – Contenders of a link are its adjacent links. University of California 21 PAMA Illustrated: a 13 5 b 23 i 7 c c k 14 21 11 f c 51 g 1. Directional links 2. Only one direction shown for simplicity 3. Hidden terminal avoidance: link (i,k) and (f,g) assigned the same code — compare node priorities of i and f. University of California 22 Summary — Unified Algorithm: • Determine the entity type (node/link); • Find out the contender set; • Run NCR to determine if the entity is active in the current time slot; • Resolve hidden terminal problem. University of California 23 Performance (Delay — Fully Connected): University of California 24 Performance (Delay — Multi-hop Network): University of California 25 Performance (Throughput — Fully Connected) University of California 26 Performance (Throughput — Multi-hop) University of California 27 Comparison with Static Scheduling Algorithm (UxDMA): University of California 28 Coloring Efficiency Comparison with UxDMA: University of California 29 Problems with NAMA • Inefficient activation in certain scenarios. – For example, only one node, a, can be activated according NAMA, although several other opportunities exist. 4 10 8 a b 1 f 7 g 6 c 5 d e 3 h —— We want to activate g and d as well. University of California 30 Node + Link (Hybrid) Activation • Additional assumption – Radio tranceiver is capable of code division channelization (DSSS —— direct sequence spread spectrum) – Code set is C . • Code assignment for each node is per time slot: i .code = i .prio mod |C | University of California 31 Hybrid Activation Multiple Access (HAMA) • Node state classification per time slot according to their priorities. – Receiver (Rx): intermediate prio among onehop neighbors. – Drain (DRx): lowest prio amongst one-hop. – BTx: highest prio among two-hop. – UTx: highest prio among one-hop. – DTx: highest prio among the one-hop of a drain. University of California 32 HAMA (cont.) • Transmission schedules: – BTx —> all one-hop neighbors. – UTx —> selected one-hops, which are in Rx state, and the UTx has the highest prio among the one-hop neighbors of the receiver. – DTx —> Drains (DRx), and the DTx has the highest prio among the one-hops of the DRx. University of California 33 HAMA Operations • Suppose no conflict in code assignment. • Nodal states are denoted beside each node: – Node D converted from Rx to DTx. – Benefit: one-activation in NAMA to four possible activations in HAMA. 10-BTx a 8-Rx b 1-DRx f 7-UTx g 6-Rx c 5-DTx d University of California 4-DRx e 3-DRx h 34 Neighbor Protocol (Need) • Purpose: propagate neighbor updates. • Cannot be based on NCR — requires a priori neighbor information. • Only way: – – – – Random access. Broadcast. No acknowledgement: why? Efficiency, broadcast. Use retransmission to improve reliability. • Why not TSMA: Topology-dependent. University of California 35 Neighbor Protocol (Method) • Insert random access section after ROMA. • Send short signals carrying neighbor updates (256 bytes). • Problem formulation: – How to regulate interval t and number n of retransmissions to have low latency to deliver messages with given (high) probability p . University of California 36 Neighbor Protocol (Results) – Reliability: deliver-probability p =99%. – Retransmission interval: t =1.44N — only depends on N (the number of two hop neighbors). – Number of retransmission: n =6.7≈7 — only depends on p . – Suppose 2Mbps bandwidth, 2 second delay, 20 two-hop neighbors — random access sections cost 9.6% of the channel resource. University of California 37 Performance Analysis • Modeling – Infinite plane with node density ρ (100 nodes per 1000mX1000m area). – Transmission range r (0m~500m). • Derive average per-node throughput according to node-distribution and node geometric relations. • Analyze both NAMA and HAMA. University of California 38 Comparison between NAMA and HAMA • HAMA has higher throughput than NAMA: – Similar at low transmission range r . – 3-4 times higher throughput at higher r . University of California 39 Comparison with CSMA and CSMA/CA (1) • Throughput of CSMA (CA) taken from the work of Yu et al. [ICNP’02]. • Load conversion: – CSMA (CA) always fully loaded. Differ at channel access probability p’ and size ldata. – HAMA load depends on packet arrival rate λ λ=p’ · ldata /(1+p’ · ldata ) • Compare the throughput S in the one-hop neighborhood N= ρπr² (ρ: node density; r Tx range). University of California 40 Comparison with CSMA and CSMA/CA (2) • Two scenarios: long data packet (100 time slots) and short data packet (10 time slot) • Different contention levels in each scenario. University of California 41 Comparison with CSMA and CSMA/CA (3) • HAMA gives the constant S at high load, whereas CSMA and CSMA/CA degrade. • HAMA differs by the shift reaching the highest S. • When the data packet is shorter, the collision vulnerable period becomes longer relatively in CSMA and CSMA/CA, thus lower throughput. University of California 42 Comparison with NAMA and UxDMA through Simulations • UxDMA schedules broadcast only, like NAMA does. • Network generated by placing 100 nodes in 1000mX1000m area. No movement. • Transmission range: 100m, 200m, 300m, 400m. • Code set size |C |=30. • Simulation duration: 100,000 time slots. University of California 43 Throughput (1) University of California 44 Throughput (2) • HAMA collected throughput of broadcast and unicast traffics separately. • Overall throughput of HAMA and NAMA is compared with the theoretical analyses — matches well. • NAMA is worse than UxDMA sometimes, HAMA is always better than UxDMA. University of California 45 Delay University of California 46 Delay Explained • UxDMA always has lower delay. • HAMA has separate delay attributes for unicast and broadcast, because they are transmitted using separate transmission opportunities. • NAMA and HAMA have the same broadcast delay. University of California 47 Conclusions: • • • • Collision-free scheduling algorithm; Minimum topology information needed; Better throughput than static scheduling algorithms. More activation opportunities can be explored in NAMA —— HAMA. • HAMA needs code division channelization. • Theoretical analyses reveal higher throughput in HAMA than in NAMA. • Scheduled approach gives higher throughput than random access approach (CSMA, CSMA/CA). University of California 48