Distinguishing Between Channel Errors and Collisions in IEEE 802.11 Muhammad Naveed Aman and Biplab Sikdar Department of ECSE, Rensselaer Polytechnic Institute, Troy, NY 12180 USA
Abstract—An important factor compromising the efficiency of IEEE 802.11 based networks and the rate adaptation algorithms they use is their inability to distinguish between packet losses caused by collisions and packet losses caused by channel errors. In this paper, we present a mechanism based on Error Vector Magnitude (EVM) to discern random channel errors from collisions in wireless networks. We show that EVM can be used not only for collision detection but it can also be used to locate the portion of a packet affected by a collision. The proposed methodology uses a classification and regression trees based procedure to develop a thresholding mechanism for EVM to distinguish between the two types of packet corruptions.
I. I NTRODUCTION The absence of infrastructure in a wireless network makes it a prime candidate for local connectivity. Carrier-Sense Multiple Access (CSMA) has been used as the main mechanism since the 1970s for sharing a common communication medium among a number of hosts. CSMA is used in both wireless (IEEE 802.11) and wired networks (IEEE 802.3). CarrierSense Multiple Access with Collision Detection (CSMA/CD) is used to improve CSMA performance by terminating a transmission as soon as a collision is detected and randomizing the retransmission time, thus reducing the probability of a second collision on retry. “Collision Detection” in wired networks means that a node listens as it transmits and can therefore detect when a frame it is transmitting has interfered (collided) with a frame transmitted by another node. In the case of a wireless network, the wireless medium makes collision detection a challenge. Firstly, the physical layer and hardware requirements make it difficult to transmit and receive at the same time. An additional complication for wireless networks is that, while a node on a wired Ethernet (IEEE 802.3) receives every other node’s transmissions, a node on 802.11 network may be too far from certain other nodes to receive their transmissions (and vice versa) [7]. Thus, a node in IEEE 802.11 is not able to directly infer if a collision has occurred and relies on other mechanisms to infer the occurrence of a collision. IEEE 802.11 approximates the behavior of collision detection by a mechanism called Multiple Access with Collision Avoidance (MACA). In MACA, the receiver sends an acknowledgment (ACK) to the sender after successfully receiving a frame, and all nodes must wait for this ACK before trying to transmit. If a timeout occurs on an ACK, the transmitter assumes that a collision has occurred. However, a
packet loss, specially in wireless environments, may also be due to channel errors attributed to weak signals and multipath characteristics of the environment. The ability to distinguish between losses due to collisions and channel errors is also useful for data-rate adaptation algorithms. Aggressive rate adaptation algorithms like AMRR and SampleRate [1], attempt to increase the throughput by using higher data rates with higher loss rates [10]. Thus, if the signal to noise ratio (SNR) is low, a packet modulated at a high data rate may be corrupted by the environment to the extent that the receiver drops it. IEEE 802.11 infers a packet loss as an indication of a collision and thus performs an exponential backoff (EB). The EB costs the transmitting node a significant amount of throughput in terms of waiting and retransmission attempts. After a number of failed retransmission attempts the cause of a packet loss is attributed to a weak signal and the rate adaptation algorithm is invoked. Therefore, understanding the root cause of a packet loss is an important factor for improving the performance of wireless networks. The correct classification of packet errors can increase the throughput by 20-60% and reduce retransmissions by 40% depending upon the channel conditions [2]. This paper addresses the problem of developing a mechanism to accurately determine if a packet loss was caused by a collision or due to channel errors. A mechanism for detecting a collision called CARA has been proposed in [4], which is based on the use of multiple RTS/CTS packets. RRAA [3] uses the CARA based RTS/CTS scheme to infer whether a packet loss is due to a collision or weak signal. A method to isolate physical packet errors from collision packet errors using RTS/CTS and packet fragmentation is given in [5]. These approaches require the observation and transmission of multiple RTS/CTS packets, thus requiring a long time to isolate the cause of a packet loss. In contrast, our approach to collision detection is more direct and is based on a metric that can be obtained immediately from the received packet, thus giving us immediate results in real time. In this paper we propose a mechanism based on the use of error vector magnitudes, and classification and regression trees to determine the root cause of a packet loss. The proposed methodology to address the issue of collision detection works by finding a threshold for the EVM to distinguish between random channel errors and collisions. The threshold is found using a classification and regression tree (CART). Using extensive simulations we show that the EVM data gathered
TABLE I A LLOWED EVM VALUES VERSUS DATA RATE IN IEEE 802.11
Fig. 1.
Data rate (Mbits/s)
EVM (dB)
EVM (%)
6 9 12 18 24 36 48 54
-5 -8 -10 -13 -16 -19 -22 -25
56 40 32 22 16 11 8 6
Illustration of Error Vector
can be used to create a very accurate CART model, to predict collisions from the EVM value of a packet. The rest of the paper is organized as follows. In Section II we present an overview of EVM. In section III we describe our collision detection mechanism. Section IV presents the simulation results. Section V presents the related work and Section VI concludes the paper. II. BACKGROUND This section provides a brief overview of EVM. An error vector is the difference between the complex voltage value of an ideal symbol and the actual received symbol. The rootmean-squared value of the error vector is defined as EVM. If In denotes the reference or transmitted signal and Rn denotes the received (distorted) signal, then Figure 1 shows the error vector En = Rn − In . Then the EVM is defined as [12]: s s PT −1 PT −1 1 1 2 2 |R − I | n n n=0 n=0 |En | T T = EV MRM S = P0 P0 (1) where P0 is the average power of all the symbols for a given modulation, and T is the number of received symbols. P0 normalizes the EVM so that it does not depend on the modulation order. We will consider wireless networks that use orthogonal frequency division multiplexing (OFDM). Let in,k denote the nth time domain OFDM symbol over sub-carrier k, then in,k can be obtained from In,k , the nth transmit symbol at the kth subcarrier, as [8]. in,k = IDF T {In,k } =
N −1 X
Fig. 2.
Scenario for introducing collisions
experience various distortions during transmission, resulting in high values of EVM. The standards specify limits on the EVM to ensure satisfactory in-band performance [6]. Table I shows the EVM thresholds specified in the IEEE 802.11 standard [11]. III. A C OLLISION D ETECTION M ECHANISM FOR IEEE 802.11 In this section we present out methodology for distinguishing between collision and channel error related packet errors. Consider the scenario in Figure 2 and assume an OFDM system with N sub-carriers. For simplicity of notation we consider BPSK modulation. Let us denote a BPSK data symbol as in,k and assume a multipath fading channel. Then rn,k = Hn in,k + ηn,k + ζn,k
where Hn is the channel coefficient in frequency domain, in,k and rn,k are the transmitted and received symbols respectively, and ηn,k and ζn,k are the additive white Gaussian noise and interference (due to a collision) over sub-carrier n and in OFDM symbol k. For the case where there is no collision, the error vector will be given by en,k = in,k − rn,k = in,k − (Hn in,k + ηn,k )
In,k ej2πkn/N
(2)
k=0
where k = 0, 1, · · · , N − 1 for N subcarriers, and n = 0, 1, · · · , T − 1. Applying Parseval’s theorem, we can rewrite (1) as s PT −1 PN −1 2 n=0 k=0 |en,k | (3) EV MRM S = T · N · P0 where en,k = rn,k − in,k , and rn,k = IDF T {Rn,k }. Equation (3), gives us the EVM value for a packet containing T OFDM symbols. The reference OFDM signal in,k may
(4)
(5)
Let us assume that a colliding packet starts in the middle of a received packet, at OFDM symbol n0 . In this case the error vector can be written as in,k − Hn in,k − ηn,k n < n0 en,k = (6) in,k − Hn in,k − ηn,k − ζn,k n ≥ n0 This shows that the absolute value of the error vector for OFDM symbols located after the start of the collision will be higher as compared to the OFDM symbols located before the collision. Consequently the EVM values for packets involved in a collision will be higher than the EVM value for
Fig. 3.
CDF of Error Vector Magnitude
packets that are lost due to channel errors. Figure 3 shows a comparison of the cumulative distribution function (CDF) for the two causes of packet losses. For clarity we only show data rates of 24, and 48 Mbps. We can see from this figure that the EVM for the packets involved in a collision is much higher than the collision free packets. We can also see that the EVM is not data rate dependent, which gives it a very good advantage since an OFDM symbol in IEEE 802.11 may contain symbols modulated using different modulation orders. From the preceding discussion we can deduce that there is a clear distinction in the distribution of EVM for the two categories. Our approach is to find a threshold γ, and on the basis of this γ the receiver carries out actions that determine the root cause of the packet loss. The sequence of actions followed by the receiver for this classification is shown in the flow diagram of Figure 4. Figure 4 shows that by detecting the root cause of a packet loss, we can invoke the correct procedure i.e. exponential backoff or rate adaptation, for subsequent transmissions. Once a collision is detected we can also find the location of the bits involved in the collision by looking at the EVM values of individual OFDM symbols. The symbols which have high EVM values will be the ones that will be in error and need to be retransmitted. Using this scheme the receiver can not only detect collisions but also detect the location of a collision and based on that the receiver can ask the transmitter to retransmit only that portion of a packet that was corrupted by the collision. Finally we note that a classification tree was used to find the threshold γ. Using simulation data, a classification and regression tree model was trained. The objective of the model is to find a set of rules which can be used to determine the value of a dependent variable Y (denoting if a packet loss was caused by a collision or a channel errors) from known values of an explanatory variable X (EVM values). From our simulation we provide a set of initial data for X where the cause of packet loss (i.e. Y ) is labeled. We then build trees using this initial data and the goal of this process is to maximize the homogeneity of the values of the dependent variable Y in the
Fig. 4.
Flow Diagram for Collision Detection
various partitions. IV. S IMULATION R ESULTS In this section we present simulation results to validate the proposed methodology for detecting collisions. The simulation tool was created in MATLAB Simulink. The transmitter and receiver models were created according to the IEEE 802.11a specifications [11]. The modulation parameters dependent on the data rate are given in Table II and the timing related parameters (associated with OFDM) for the simulator are given in Table III. The simulations assume a frequency flat multipath Rayleigh channel using the Jake’s model [9]. To increase the probability of packet loss due to weak signal, we use a maximum Doppler frequency of 100Hz, thereby modeling the effect of highly mobile users within the network. We only consider data rates of 12Mbps (QPSK), 24Mbps (16QAM), and 48Mbps (64QAM) for the purpose of our simulations. The network topology used in the simulations consists of two transmitters and one receiver and is shown in Figure 2. The distance between the receiver and any of the transmitters can be changed. T1 sends 32 OFDM symbols per frame, while the frame size for T2 is variable. The frame size of T2 follows a uniform distribution between 1 and 32 symbols. This allows us to simulate collisions occurring between packets
TABLE II A LLOWED EVM VALUES VERSUS DATA RATE IN IEEE 802.11 Data rate (Mbits/s)
Modulation Coding rate (R)
Coded bits per subcarrier (NBP SC )
Coded bits per OFDM symbol (NCBP S )
Data bits per OFDM symbol (NDBP S )
6 9 12 18 24 36 48 54
BPSK BPSK QPSK QPSK 16QAM 16QAM 64QAM 64QAM
1 1 2 2 4 4 6 6
48 48 96 96 192 192 288 288
24 36 48 72 96 144 192 216
1/2 3/4 1/2 3/4 1/2 3/4 2/3 3/4
Fig. 5.
Classification tree - Low Capture Effect
TABLE III T IMING RELATED PARAMETERS Parameter
Value
NSY M : Samples per OFDM symbol NF F T : FFT Length NSD : Number of data subcarriers NSP : Number of pilot subcarriers NST : Number of total subcarriers NT RAIN : Number of training symbols ∆F : Subcarrier frequency spacing TF F T : IFFT/FFT period TP REAM BLE : Preamble duration TGI : Guard Interval (GI) duration TGI2 : Training symbol GI duration TSY M : Symbol interval TLON G : Training sequence duration
80 64 48 4 52 (NSD + NSP ) 2 0.3125 MHz (=20 MHz/64) 3.2µs(1/∆F ) 8µs 0.8µs (TF F T /4) 8µs 4µs (TGI + TF F T ) 8µs (TGI2 + 2xTF F T )
having different sizes. The simulator also allows us to set the probability of collision, thus giving us the flexibility of controlled experiments. We evaluate our root cause detection mechanism on the basis of: (i) the probability of false alarm PF A - that is, the case when a channel error is miss-classified as a collision, (ii) the probability of miss-detection PM D - that is, the case when a collision is miss-classified as a channel error, and (iii) the accuracy - that is, the number of cases our mechanism classifies packets correctly. Our results are grouped in terms of the capture effect in the network. Capture effect is a phenomenon in which the transmission of the weaker signal is suppressed by the transmission of a strong signal at the receiver, where the weaker signal is attenuated. We present the results of our root cause detection mechanism by considering the following scenarios according to the capture effect. In each of the following scenarios we simulate 3910 packets (32 OFDM symbols per packet), where we have an equal number of packets from the three data rates. 1) Scenario 1: Low Capture Effect: We first consider the case where both the transmitters are placed at an equal distance from the receiver. We first run the simulation to get data that is used for training a CART model. To get the training data set we
Fig. 6. Scatter plot of packet EVM values, Low Capture Effect- training data set
simulate the scenario shown in Figure 2, with the transmitters placed at different distances ranging from 3m to 40m from the receiver. The training data obtained is used to create a classification tree model which in turn gives us a threshold γ for EVM. To validate our model we use γ to classify new cases, i.e. we simulate the scenario of Figure 2 with a different set of distance values for the transmitters to obtain a validation data set. We input the validation data set into our model to evaluate its performance. The classification tree obtained with the training data set is shown in Figure 5. A scatterplot showing the EVM values of packets received at receiver R for the training data set is shown in Figure 6. The dotted black line in Figure 6 shows the threshold obtained from the classification tree of Figure 5, the dark circles correspond to collision packets and the light circles correspond to packets not involved in a collision. Figure 6 shows that there is a clean separation between the two types packets i.e. packets with or without a collision. If
TABLE IV A LLOWED EVM VALUES VERSUS DATA RATE IN IEEE 802.11 d1 (m)
d2 (m)
PF A (%)
PM D (%)
Accuracy(%)
5 5 5 5 5 5 5
9 17 23 29 33 40 50
3 5 5 2 3 3 2
9 20 21 27 36 43 50
94 87 86 85 80 77 73
Fig. 7. Scatter plot of packet EVM values, Low Capture Effect - validation data set Fig. 8.
we set the EVM threshold for detecting collisions to -15.5dB in the training data set, we get an accuracy of approximately 97% and the probability of false alarm and miss-detection are 3.1% and 2.8%, respectively. If we use this threshold to detect collisions in our validation scenario we get an accuracy of 96.9%, a probability of false alarm of 3.4% and a probability of miss-detection of 2.6%. The scatter plot for the validation data set is shown in Figure 7. These results validate our model, and show that we get high accuracy to predict collisions in the presence of low capture effect. 2) Scenario 2: High Capture Effect: If one of the transmitters is closer to the receiver, the packet with stronger signal is received with almost no errors or a few bits in error. The problem of collision detection in this case is more complicated. In this scenario we place T1 at a distance of d1 = 5m and we move T2 away from the receiver i.e d2 > 5m, and we run the simulation for different values of d2 . If we apply the threshold found in scenario 1 to predict collisions for this data, we get the miss-classification rates given in Table IV. Table IV shows that the accuracy and PM D are adversely affected as the capture effect increases (i.e as we move T2 farther from the receiver), however the accuracy is still quite significant. The PF A is quite low for all the cases, which is very important for our case: we do not want to invoke exponential backoff unnecessarily.
Classification tree for OFDM Symbols inside a Collision Packet
of the EVM values for the OFDM symbols is shown in Figure 9. If an EVM threshold of -20.4dB is used to distinguish between corrupted OFDM symbols and error free OFDM symbols, we get 100% accurate results. These results show that once a collision is detected, we can identify the corrupted
A. Detecting the Location of Bits Corrupted by a Collision To detect the OFDM symbols corrupted by a collision, we use the simulation data described previously and we only consider the data for the packets that are involved in a collision. We then create a CART model using this data which gives us the classification tree shown in Figure 8. A scatterplot
Fig. 9.
Scatter plot of EVM values of OFDM symbols
OFDM signal with very high accuracy. V. R ELATED W ORK In [5] a scheme is described to isolate physical packet errors from collisions, which requires sending multiple RTS/CTS packets and fragmentation. Their scheme also requires a minimum packet size for detecting collisions and the calculation of packet error rates. The overhead of sending RTS/CTS, fragmentation, and the limitation on the packet size reduces the effectiveness of this approach. However, our collision detection mechanism does not require any additional overhead. In another work [2], collision detection is done by gathering statistics for the bit error rate, error per symbol (EPS) and the received signal strength (RSS). The authors have proposed an algorithm using these detectors and a cooperation between the access points. However the accuracy of their approach greatly relies on the cooperation between the access points which limits their practical applicability. The authors achieved an accuracy of about 60% and a PF A of about 2% on an average using three predictors. However if only one access point is used, their scheme gives an accuracy of 28% in the presence of high capture effect. From these results we can see that the accuracy of our proposed mechanism is much higher: we get an accuracy of about 86% and a PF A of about 3% on an average using only one predictor. Moreover the accuracy of our approach in the presence of high capture effect is about 83% on an average using only one receiver. Our approach to root cause detection uses only one predictor (EVM), which can be calculated even before a packet is completely decoded. We do not require feedback from any access point, rather the feedback comes from the receiver of the transmission. There have been efforts to use the information at the PHY layer to solve problems at the MAC layer. In [13] it is shown that by retransmitting only those bits in a packet that are likely in error we can increase the end-to-end capacity of automatic repeat request (ARQ) by 2x under moderate load. However, this approach requires custom built hardware, while our methodology can be implemented without the need for any specialized hardware. VI. C ONCLUSION This paper addresses the problem of determining if a packet loss in IEEE 802.11 networks has been caused by channel errors or due to a collision with another simultaneous transmission. Understanding the root cause behind a packet loss is necessary for the correct operation of the exponential backoff and rate adaptation algorithms used by IEEE 802.11. The proposed mechanism for root cause detection is based on evaluating and comparing the Error Vector Magnitude of received symbols against a threshold obtained using a classification and regression tree. In addition to detecting the root cause of a packet loos, our methodology can also be used to isolate the bits in the packet that were corrupted due to a collision.
R EFERENCES [1] J. Bicket, “Bit-rate selection in wireless networks,” MIT Master’s Thesis, 2005. [2] S. Rayanchu, A. Mishra, D. Agrawal, S. Saha, S. Benerjee, “Diagnosing Wireless Packet Losses in 802.11: Separating Collision from Weak Signal,” Proc. IEEE INFOCOM, Phoenix, AZ, April 2008. [3] S. H. Wong, H. Yang, L. Lu, and B. Bhargavan, “Robust Rate Adaptation in 802.11 Wireless Networks,” Proc. ACM MOBICOM, Los Angeles, CA, September 2006. [4] J. Kim, S. Kim, S. Choi, and D. Qiao, “CARA: Collision-aware rate adaptation for 802.11 WLANs,” Proc. IEEE INFOCOM, Barcelona, Spain, March 2006. [5] M. Khan, and D. Veitch, “Isolating Physical PER for Smart Rate Selection in 802.11,” Proc. IEEE INFOCOM, Rio de Jeneiro, Brazil, April 2009. [6] C. Zhao, and R.J. Baxley, “Error Vector Magnitude Analysis for OFDM Systems,” Signals, Systems and Computers, 2006. ACSSC ’06. Fortieth Asilomar Conference on , October 2006. [7] L. L. Peterson, and B. S. Davie, “Computer Networks - A systems Approach,” Elsevier Inc., 2007. [8] Y. S. Choo, J. Kim, W. Y. Yang, and C. G. Kang, “MIMO-OFDM Wireless Communications with MATLAB,” IEEE press, 2010. [9] W. C. Jakes, “Microwave Mobile Communications,” IEEE press, 1974. [10] A. U. Joshi, and P. Kulkarni, “Vehicular WiFi Access and Rate Adaptation,” Proc. ACM SIGCOMM, New Delhi, Inida, August 2010. [11] IEEE Standard 802.11a-1999, “High-speed Physical Layer in the 5 GHz Band,” 1999. [12] S. Forestier, P. Bouysse, R. Quere, A. Mallet, J. M. Nebus, and L. Lapierre, “Joint Optimization of the Power-Added Efficiency and the Error-Vector Measurement of 20-GHz pHEMT Amplifier through a New Dynamic Bias-Control Method,” IEEE Transactions on Microwave Theory Tech., vol.52, no.4, pp.1132-1141, 2004. [13] K. Jamieson and H. Balakrishnan, “PPR: Partial Packet Recover for wireless networks,” Proc. ACM SIGCOMM, 2007.