XPRESS: A Cross-Layer Backpressure Architecture for Wireless Multi-Hop Networks

1. Rafael Laufer, TheodorosSalonidis, Henrik Lundgren, Pascal Le Guyadec, MobiCom2011 XPRESS: A Cross-Layer Backpressure Architecture forWireless Multi-Hop Networks 2011/12/19study group by Neight

2. Motivation XPRESS Backpressure Scheduling &Routing Cross-Layer Protocol Stack ContorlPlane Congestion control Optimal Schedule Computation OutLine

3. Motivation • Wireless multi-hop networks operate below capacity • Poor coordination across layers • Poor coordination among transmitting nodes • How to achieve the network capacity?

4. A radical alternative by advocating cooperation among the multiple layers of the protocol stack algorithm which, in theory, achieves the network capacity NEW IDEA

5. NEW IDEA • Backpressure scheduling & Routing • Cross-Layer cooperation • Select optimal link set for transmission • TDMA MAC protocol

6. XPRESS NEW IDEA

7. XPRESS view Internet MC GW MAP CS DS … … … … Frame k • Mesh access point (MAP) • Sends queue lengths • Executes the schedule • Cross-layer protocol stack • Mesh controller (MC) • Receives flow queue lengths • Computes schedule • Disseminates schedule

8. Challenges • Practical challenges • Time slots: TDMA MAC in multi-hop networks • Link sets: Knowledge of non-interfering links • Protocol overhead: Queue backlogs known at each slot • Computation overhead: Exhaustive search over links sets • Link scheduling: Backpressure schedules links, not nodes • Hardware constraints: Memory limitations at wireless cards • Backpressure so far a theoretical concept • No real system implementing backpressure to date

9. Backpressure Scheduling Algorithm

10. Backpressure Scheduling & Routing • Flow schedule and Routing • For each link ( i,j ), select the flow with the maximum queue differential backlog • Link Schedule • Compute the weight of link as • Select links to maximize • Transmit chosen flows on the selected links

11. Backpressure Scheduling & Routing 6 2 3 5 3 B A 2 1 6 3 4 5 C D 4 7 7 8 1 • Transmission: • During the time slot, a selected link ( i,j) transmits a packet from flow • using rate

12. Backpressure Scheduler Implement

13. Backpressure Scheduler • Challenge: compute optimalschedule per slot • Knowledge of queue backlogs at each slot • Speculative scheduling: estimate queue backlogs • Challenge: schedule computation takes time • During frame , compute the schedule for frame Estimate Estimate MC Compute Compute Execution of Execution of CS DS CS DS MAP Frame Frame

14. Cross-Layer cooperation Implement

15. XPRESS Cross-Layer Protocol Stack Link Schedule Flow Schedule Flow Classifier Link Classifier PreQ FlowQ LinkQ A1 Time . . . An Link scheduler enforces schedule, respecting TDMA slot boundaries Per-Flow Queues Packet Scheduler Congestion Control Packet Scheduler Flow queues at the kernel address the limited memory in the firmware Per-Link Queues Slot t+1 Congestion control ensures flow rates are within the capacity region Flow scheduler enforces schedule and avoids overflows at the firmware Forward Link queues required for link scheduling Local User Kernel Firmware Wireless

16. XPRESS flows: • General and can easily accommodate other flow definitions. • compared to the usual 5-tuple flow definition of source and destination IP addresses, source and destination transport ports, and transport protocol (i.e., TCP or UDP), • this design decision reduces processing and communication overhead in XPRESS at the expense of flow granularity • Congestion control and flow scheduling: • Each flow has two individual queues, namely, a pre-queue (PreQ) and a flow queue (FlowQ) • Congestion control is performed according to and depends only on the length of the local FlowQ. • A longer FlowQ reduces the allowed input rate, while a shorter FlowQ allows a higher rate. XPRESS Cross-Layer Protocol Stack

17. Link scheduling: • The MAC protocol keeps an individual queue for each neighbor in order to enable link scheduling, which allows a higher spatial reuse than node scheduling. • The slotted MAC, realized by a TDMA MAC protocol maintains network-wide node synchronization, and ensures that transmissions occur strictly within slot boundaries • Packet reception and forwarding: • Once a packet is received, it is first filtered based on the destination MAC address and then inserted into a single receive queue (RxQ) at the firmware. • The packet is delivered to the network layer at the kernel, where it is routed and tagged for local delivery or forwarding XPRESS Cross-Layer Protocol Stack

18. ContorlPlane Implement

19. Contorl Plane

20. Congestion Control Implement

21. Congestion control A simple distributed congestion control algorithm where the source s of each flowf adjusts the flow rate as where is the queue backlog for flow f at the source s and is the inverse of the first derivative of the utility function at the point In XPRESS

22. Optimal Schedule Computation More Implement problem

23. Optimal Schedule Computation  Conflict graph • For each slot, exhaustive search over all link sets • Find link set which maximizes the sum of weights • Binary interference in TDMA MAC over 802.11 PHY • Links have either low or high PDR • Maximum weighted independent set (MWIS) • MWIS computation takes 100 µs for testbed

24. A direct approach Finding the link transmission sets and their capacities isa challenge because each link capacity depends on both thechannel condition and the interference created by the otherlinks in the set. A direct approach requires O ( N r × 2L)

25. Binary interferenceexperiment Define: • The capacity μijof each link ( i,j ) is estimatedon a TDMA frame time scale as Pij× Rij • Pijis thepacket delivery ratio (PDR) and • Rijis the PHY rate of link( i,j ) during the TDMAframe.

26. Binary interferenceexperiment In order to understand how interference manifests in ourTDMA system, we perform experimentfor alllink pairs that do not share a node in testbed The linksof each pair simultaneously transmit backlogged UDP trafficfor 1 minute using broadcast packets at 24 Mbps PHY rate.During this time, the receivers measure the received signalstrength (RSS) and PDR values at each TDMA frame

27. Binary interference

28. In Backpress scheduler: • Binary interference  MWIS maximum weight independent set (MWIS) computation in system due tobinary interference. • The MWIS computation is an NP-hard implementation is based on an algorithm for enumerating maximal independent sets at the beginning of each frame. We then find the MWIS using a linear search over the independent sets. • Heap structure For efficiency, these sets could be stored in a heap structure keyed by their weights. At each slot, queue lengths change, which triggers a heap update. After the update, the new MWIS can be found as the root of the heap More Implement problem

29. MWIS http://www.csie.ntnu.edu.tw/~u91029/Matching.html

30. Use RSS measurement: This results in one RSS measurement per framefor each link. Based on these measurements, we estimatethe signal-to-interference ratio (SIR) of a link ( i,j ) underthe interference of a transmitting node k as the differenceS ij − S kj , between S ij , the RSS of link ( i,j ), and S kj , theRSS of link ( k,j ), both in dBm. If the SIR of the link exceeds a threshold, which depends on the PHY rate, the linkPDR is estimated“high.” GENERATE Conflict graph

31. RSS BASE PROBLEM: • RSS values reported by 802.11 wireless cards can be highly inaccurate due to the type of hardware, poor calibration, environmental conditions, location, temperature, multi-path effects, and external interference. • it relies on the RSS of decoded packets and hence cannot detect hidden interferers which are within interference range, but not within communication range. • Solution: • use PDR measurements GENERATE Conflict graph

32. The MC uses a conflict graph to represent interference in the network. A vertex vijin the conflict graph corresponds to the link ( i,j ) in the network graph. An edge between vertices vijand vkldenotes interference between links ( i,j) and ( k,l ) in either the DATA or ACK directions. The vertex independent sets in the conflict graph correspond to the link transmission sets. The conflict graph update mechanism is executed at each TDMA frame, after the MC receives the RSS and PDR measurements. The conflict graph construction occurs in two stages. IMPLEMENTConflict graph

33. Conflict graph

34. If Sijis not measured, the transmission from i to j is estimated outside of communication and interference range. MC creates a vertex vijin the conflict graph if S ij ≥ the RX sensitivity threshold of receiver j at PHY rate R . MC adds an edge between each pair of vertices vijand vklin the conflict graph if either they share a common nodeor if the SIR of DATA or ACK directions is less than the SIR receiver threshold at PHY rate R STAGE1

35. For each link ( i,j ) selected by the first stage, the MC checks its reported PDR value Pij. If Pij≥ 90%, the link remains in the conflict graph. Otherwise,MC finds hidden interferers. The MC identifies the hidden interferers using the connectivity graph. we assume that the hidden interferers of link ( i,j ) are those transmitters in Iijwhich are two-hopneighbors of either i or j . For each such node k , the algorithm adds an edge between vertex vijand vertex vklin the conflict graph. If link ( i,j ) fails again in the same set Iij during the next TDMA frame, the three-hop neighbors of i and j can be considered, until the hidden interferers are detected. STAGE2

36. Like Four Color Theorem http://www.csie.ntnu.edu.tw/~u91029/Coloring.html

37. IF NEED some REFERENCE http://caterpillar.onlyfun.net/Gossip/AlgorithmGossip/HeapSort.htm

38. Interference Estimation • Knowledge of interference to build conflict graph • Naive approach: measure each link set at all rates • Measurement complexity • RSS measurements taken on each TDMA frame • Control packets used to measure RSS • Link RSS used to compute SIR  threshold per PHY rate • Measurement complexity reduced to • RSS limited only to decoded packets • PDR measurements also taken on each TDMA frame • Detection of hidden interferers

39. Other ISSUES • Link queue: • Head-of-blocking • This problem may occur if we have wireless losses and the packet at the head of the queue is destined for a different neighbor than the one assigned for the slot, resulting in no packet transmitted during that slot.

40. Head-of-blocking • 當封包流入的總和比交換switch的速度快封包會被儲存在input queue等待交換。 • Head-of-the-Line (HOL) blocking: 在等待交換的封包影響後來的封包進行交換 (queue的特性)。 擷取自計算機網路概論 課程投影片

41. Contributions • Design and implementation of XPRESS • First throughput-optimal backpressure system • Backpressure challenges addressed • Time slots: multi-hop TDMA MAC & time synchronization • Link sets: RSS-based interference estimation • Protocol overhead: Multi-slot framing and speculation • Computation overhead: Binary interference  MWIS • Link scheduling: Individual link queues at the MAC • Hardware constraints: Network/MAC queue coordination

42. 802.11a Indoor Testbed • MAP node • 1.6 GHz CPU, 512 MB RAM • Linux OS / BP kernel module • 802.11 Technicolor card (5 GHz) • Customized firmware (TDMA/link scheduling) • Mesh controller • 2.7 GHz CPU, 16 GB RAM

43. Multi-Hop: Multi-Path Topology http://nccur.lib.nccu.edu.tw/bitstream/140.119/32668/6/97101206.pdf • Ability of XPRESS to exploit multiple paths • One flow between extreme nodes • XPRESS allowed to use every link available • 802.11 uses the shortest ETX path

44. Multi-Hop Throughput test Use Iperf

45. Multi-Hop: Multi-Path Topology Coordination & path diversity  higher network throughput

46. Multi-Hop Select short path (less hops) Delay is linear increase after meet throughput

47. Queue Backlog Estimation Error Accurate predictions  XPRESS reaches network capacity

48. Overhead: Computation • MWIS computation for optimal schedules • In theory, MWIS is NP-hard • In practice, polynomial with the number of links

49. Overhead: Computation MWIS computation is feasible for practical network sizes

50. Overhead: Protocol • Each frame • Queue backlogs sent from the MAPs to the MC • Computed schedule sent from the MC to MAPs • Time to exchange this on the control subframe