Presentation on theme: "Doc.: IEEE 802.11-11/0953r1 Submission July 2011 Edward Reuss, SK CommunicationsSlide 1 Channel Contention with a Large Number of Devices Date: 2011-07-14."— Presentation transcript:
doc.: IEEE /0953r1 Submission July 2011 Edward Reuss, SK CommunicationsSlide 1 Channel Contention with a Large Number of Devices Date: Authors:
doc.: IEEE /0953r1 Submission July 2011 Edward Reuss, SK CommunicationsSlide 2 Abstract Certain use cases for IEEE networks involve a large number of devices conveying time-sensitive information through a single AP. Many of these involve frequent transmission of relatively short packets, less than 200 bytes. In these scenarios the time on the medium to transmit the actual data is very short, but because there are so many devices the contention windows must be very long to minimize the probability of a collision. This effect limits the maximum number of devices that can participate in these applications. This presentation introduces the reasons that cause this problem and proposes some possible solutions.
doc.: IEEE /0953r1 Submission July 2011 Edward Reuss, SK CommunicationsSlide 3 Channel Contention with a Large Number of Devices Several applications require support for a large number of devices communicating time-sensitive information, such as telephony or sensor streams, through a single AP. Most of these use cases involve frequently transmitting a series of relatively short data packets (~200 bytes or less). –Enterprise Wireless VoIP –Internet of Things (IoT) –SmartGrid –Process Control (Manufacturing, Industrial systems, etc.) –Automotive (Engine sensors, Impact warnings, etc.) –??
doc.: IEEE /0953r1 Submission July 2011 Edward Reuss, SK CommunicationsSlide 4 Example: Enterprise Wireless VoIP Assume: –Either: Duplex G.711 PCM 8 kSamples/sec. or Duplex G.722 ADPCM 16 kSamples/sec. –20 msec. packet intervals = 160 bytes of audio data per packet, uplink and downlink –UDP/IP Packets: 86 header bytes total SNAP/LLC: 8 bytes, UDP header: 8 bytes, IP header: 20 bytes MAC Security (AES): 16 bytes, MAC frame header & CRC: 34 bytes –IEEE a PHY: Symbol Interval: 4 μsec, PHY header: 16 μsec, PLCP header: 4 μsec.
doc.: IEEE /0953r1 Submission Example: Enterprise Wireless VoIP #2 Calculate the length of a single audio data packet + the Ack packet + DIFS + SIFS, in microseconds. Therefore the maximum theoretical call capacity for duplex data packets, Ack packets, DIFS & SIFS is: July 2011 Edward Reuss, SK CommunicationsSlide 5 IEEE a TX+Ack+SIFS +DIFSDuty CycleEff. # Devices 6 Mbps % Mbps % Mbps % Mbps % Mbps % Mbps % Mbps % Mbps %70.42
doc.: IEEE /0953r1 Submission IEEE Collision Avoidance Contention Window –New Packets: Randomly assigned to the range [0,CWmin]. –Retransmitted Packets: Randomly assigned to a range that exponentially increases from [0,CWmin] to [0, CWmax]. Assuming no collisions & no retransmissions –For k devices attempting to transmit a packet on the same Access Category, the average contention window delay is: July 2011 Edward Reuss, SK CommunicationsSlide 6
doc.: IEEE /0953r1 Submission Effect of CA on the VoIP Example Add t CWavg to each audio data transmission –Assumes 20 devices and CWmin = 127. Times are in microseconds. This overhead remains tolerable, although the average length of the contention window is approaching the length of the entire packet + Ack sequence. July 2011 Edward Reuss, SK CommunicationsSlide 7 IEEE a TX+Ack+SIF S+PIFSt CWavg Duty Cycle Eff. # Devices 6 Mbps % Mbps % Mbps % Mbps % Mbps % Mbps % Mbps % Mbps %50.21
doc.: IEEE /0953r1 Submission Probability of a Collision Pigeonhole Principle: –If k pigeons are placed randomly into m pigeonholes (m k), the probability of 2 or more pigeons in one hole are: –Basis for the Birthday Paradox How many randomly chosen people do you need in a room for there to be a better than even chance that at least two of them have the same birthday? Answer: 23 people The paradox is that this is a surprisingly small number for 365 days in a year. –This is the source of the problem with IEEE CA for many devices transmitting rapid sequences of short packets. July 2011 Edward Reuss, SK CommunicationsSlide 8
doc.: IEEE /0953r1 Submission Probability of a Collision - Example July 2011 Edward Reuss, SK CommunicationsSlide 9 Cwmin # of Devices "k" Prob. Collisiont CWavg μsec. # of Devices "k" Prob. Collisiont CWavg μsec Assume 8 & 20 devices for CWmin = 31 to TS = 9 μsec. To reduce the probability of collisions, CWmin must be very large, causing the average contention window delay (t CWavg ) to grow. – t CWavg becomes larger than the audio packet exchange time (462 to 142 μsec), dramatically reducing the maximum packet throughput even further. –The impact of retransmitting the collided packets will reduce the maximum packet throughput capacity even further. –*Note: The listed t CWavg for 8 devices is an artifact of 9 μsec / (8+1) = 1.
doc.: IEEE /0953r1 Submission Impact of Collisions - Cascade Every time a collision occurs, at least two packets must be retransmitted. At the same time another device may have a packet ready to transmit. Therefore, for the next contention window: k = k + 1 If CWmin = CWmax = 255, there is no room for the exponential backoff for the retransmitted packets, so the probability continues to rise. This causes the probability of a second collision to increase, resulting in a cascade of packet collisions and retransmissions. July 2011 Edward Reuss, SK CommunicationsSlide 10
doc.: IEEE /0953r1 Submission Future Trends Amendments IEEE n and ac exacerbate this problem further by increasing the relative overhead due to the packet headers and the contention window with respect to the actual user data symbols. –IEEE a may use as few as 10 symbols (40 μsec.) for the VoIP example. –IEEE ac can fit an entire VoIP data packet into a single symbol at the higher modulation indices. –The utilization efficiency of the medium falls as the contention window and packet header overheads dominate the medium. July 2011 Edward Reuss, SK CommunicationsSlide 11
doc.: IEEE /0953r1 Submission Impact on Short Packet Applications These amendments have increased the effective throughput for long packets from a single stream by about 40 times, from about 25 Mbps for a to over 1000 Mbps for ac. But the contention window remains unchanged from the original definition in IEEE and the Time_Slot of 9 μsec, as defined in a. This limits the utility of IEEE for short packet applications. July 2011 Edward Reuss, SK CommunicationsSlide 12
doc.: IEEE /0953r1 Submission Solution 1: – Fractional Time Slots Add an optional operating mode that recognizes fractional time slots FTS. –FTS = TS / 4 = 9 μsec / 4 = 2.25 μsec. Or possibly: –FTS = TS / 8 = 9 μsec / 8 = μsec. This is a cheap solution, but it may be impractical to implement. This is only a partial solution, but it may be adequate to make the use cases feasible. July 2011 Edward Reuss, SK CommunicationsSlide 13
doc.: IEEE /0953r1 Submission Solution 2: Dynamic Contention Window Modify the contention window dynamically according to the load requirements. Adapts to high load conditions when required. Can choose between various strategies. –Cooperative or non-cooperative Cooperative: All nodes share their load requirements with all other nodes. Non-cooperative requires each node to listen to the traffic, collisions, etc, to infer the network load requirements. –Heterogeneous or homogeneous Homogeneous uses the same strategy at all nodes. Heterogeneous uses different strategies for differentiated access categories. –Reference  uses a game theoretic strategy amongst non-cooperative nodes. Must guard against cheaters. July 2011 Edward Reuss, SK CommunicationsSlide 14
doc.: IEEE /0953r1 Submission Solutions – Contention-less Strategies Removes the contention window entirely. –See reference . Theoretically removes all collisions. (Famous last words) Normally cooperative (but not required) –Every node broadcasts, or group-casts, their load requirements to all of the other nodes. This overhead is much smaller than the overhead from the contention windows and the collisions. –All nodes use an identical strategy to calculate the transmission schedule for every node in the group. (Homogeneous strategy) –Each node knows when it is their turn to transmit their packet(s) according to the calculated schedule. July 2011 Edward Reuss, SK CommunicationsSlide 15
doc.: IEEE /0953r1 Submission Contention-less Strategy in IEEE AP sends CTS-to-Self to set the NAV, defining the contention-less interval according to the dynamic load requirements for each interval. Each device senses the contention-less interval, and starts transmitting at their calculated point in the schedule. Similar to the old CFP, except more flexible. The load announcements can be piggy-backed onto other existing traffic, such as sounding packets. July 2011 Edward Reuss, SK CommunicationsSlide 16
doc.: IEEE /0953r1 Submission Straw Poll Should Channel Contention with a Large Number of Devices be pursued in IEEE ? Result –Agree: 35 –Disagree: 0 –Abstain: 25 July 2011 Edward Reuss, SK CommunicationsSlide 17
doc.: IEEE /0953r1 Submission July 2011 Edward Reuss, SK CommunicationsSlide 18 References 1.IEEE Amendment IEEE a Amendment IEEE n 4.Draft Amendment IEEE ac 5.Contention Control: A Game-Theoretic Approach, Chen, Low & Doyle, 46 th IEEE Conference on Decision and Control, Many-to-many communication for mobile ad hoc networks, Moraes, Sadjadpour & Garcia-Luna- Aceves, IEEE Transactions on Wireless Communications, Vol. 8 Issue 5, May 2009.