ChASER: Channel-Aware Symbol Error Reductionfor High-Performance WiFi Systems in

Dynamic Channel EnvironmentOkhwan Lee†, Weiping Sun†, Jihoon Kim†, Hyuk Lee†, Bo Ryu‡, Jungwoo Lee† and Sunghyun Choi†

†Department of ECE and INMC, Seoul National University, Seoul, 151-744, Korea‡EpiSys Science, Compton, CA 90220, U.S.A

{ohlee, weiping, jhkim}@mwnl.snu.ac.kr, hyuklee@wspl.snu.ac.kr, boryu@episyscience.com, {junglee, schoi}@snu.ac.kr

Abstract—Due to considerable increases in user mobility andframe length through aggregation, the wireless channel remainsno longer time-invariant during the (aggregated) frame trans-mission time. However, the existing IEEE 802.11 standards stilldefine the channel estimation to be performed only once at thepreamble for coherent OFDM receivers, and the same channelinformation to be used throughout the entire (aggregated) frameprocessing. Our experimental results reveal that this baselinechannel estimation approach seriously deteriorates the WiFiperformance, especially for pedestrian mobile users and therecently adopted frame aggregation scheme.

In this paper, we propose Channel-Aware Symbol ErrorReduction (ChASER), a new practical channel estimation andtracking scheme for WiFi receivers. ChASER utilizes the re-encoding and re-modulation of the received data symbol to keepup with the wireless channel dynamics at the granularity ofOFDM symbols. Our extensive, trace-driven link-level simulationshows significant performance gains over a wide range of channelconditions based on the real wireless channel traces collectedby the off-the-shelf WiFi device. In addition, the feasibility ofits low-complexity and standard compliance is demonstratedby Microsoft’s Software Radio (Sora) prototype implementationand experimentation. To our knowledge, ChASER is the firstIEEE 802.11n-compatible channel tracking algorithm since otherapproaches addressing the time-varying channel conditions overa single (aggregated) frame duration require costly modificationsof the IEEE 802.11n standard.

I. INTRODUCTION

OVER the last quarter century, the high-speed wirelesslocal area network (WLAN) technology has under-

gone tremendous growth since its inception, offering severalhundred-fold increases in data rate and ubiquitous access at ex-tremely low cost. Much of this technological and commercialsuccess is attributed to WiFi, the hallmark of the IEEE 802.11WLAN working group (WG), as it is now an indispensablepart of our daily lives with the ever-increasing demands forhigher data rates [1].

Since it was first introduced, WiFi technology has beenprimarily designed for indoor and nomadic (but stationaryduring communication) users where the assumption of quasi-stationary channel over the duration of a single frame trans-mission has matched the reality extremely well. When thewireless channel does not change significantly during theperiod of a single frame transmission (i.e., frame durationis within the scale of channel coherence time), the channel

state information (CSI) needs to be estimated only once atthe preamble at the beginning of the frame, and can beexploited throughout the entire frame reception process. Theexisting IEEE 802.11 standards abide by this practice of“estimate once, apply throughout” during the receiver processof demodulation/decoding.

Two recent trends, however, are challenging this fundamen-tal assumption of the quasi-stationary channel condition. First,the usage pattern of WiFi users is increasingly becoming “mo-bile during communication,” thanks to the explosive growthof various portable devices such as smartphones, tablets,and emerging wearable gadgets. Second, the new frame-aggregation scheme, called aggregated MAC protocol/servicedata unit (A-MPDU/A-MSDU), has been adopted by recentWiFi standards such as IEEE 802.11n and IEEE 802.11ac inorder to enhance the throughput performance by minimizinginter-frame protocol overhead. We have found, through exten-sive experiments, that the increased mobility and frame lengthcaused by these trends are likely to change the channel condi-tion significantly during the single reception of an (aggregated)frame. This makes the channel response experienced by thelatter part of the frame considerably different from the oneexperienced by the preamble. In other words, the uncertaintyof the CSI obtained at the preamble increases significantlywith the increase in the frame duration and user mobility.

Motivated by the strong evidence that the changes in chan-nel conditions during a frame reception leads to poor WiFiperformance with user mobility and increased frame length, wepropose Channel-Aware Symbol Error Reduction (ChASER),a novel standard-compliant channel tracking scheme thatconstantly updates the CSI during the entire course of theframe reception process. This is done by reconstructing thetransmitted data symbols (in parallel with the normal receiverprocessing) and comparing it to the received data symbols sothat the channel response can be estimated at any part of theframe at the granularity of a data symbol. Since reconstructingtransmitted data symbols reliably (with minimum errors) andquickly (within the frame processing time) is the key enablerfor symbol-level channel estimation and tracking, ChASERwittingly reuses the exactly same modulation and codingoperations available in the transmitter’s signal processingblocks (DSP) with the addition of a low-complexity adaptive

L-STF L-LTF L-SIG

L-preamble HT-preamble A-MPDU subframeMPDU

delimiterMSDU CRC Tail Pad

HT-SIG HT-STF HT-LTF

PLCP

header

Pilot symbol

MAC

headerPad

A-MPDU subframe

Fig. 1: IEEE 802.11n MAC/PHY mixed-mode PPDU frame format of A-MPDU.

filter. The adaptive filter is required to alleviate the errorpropagation incurred by falsely reconstructed data symbols.Furthermore, no extra protocol overhead is needed, and theestimation process can be done in parallel with the decodingprocess. With a ChASER-enabled WiFi receiver, the CSIobtained from the preamble is replaced whenever a new CSIis estimated until the frame reception processing ends.

Trace-driven link-level simulation results with a large num-ber of CSI measurements from real-life experiments showlarge performance gains (as high as 99.5% improvement) overa wide range of channel conditions. We have also implementedChASER on Microsoft’s Software Radio (Sora) testbed, anddemonstrated the feasibility of easy implementation and stan-dard compliance commercial 802.11n access point. Theseresults offer sufficient evidence that the proposed ChASERenables the future WiFi devices to readily accommodate theemerging paradigm shift in increasing mobility and frameaggregation.

The rest of the paper is organized as follows. Section IIprovides a survey of related work that addresses the channelestimation within a frame duration. Then, Section III intro-duces the background and motivation of this work based on theextensive analysis of the degree of wireless channel variationmeasured in real-life testbed. Sections IV to VI present thedetails of ChASER, performance evaluation methodology,and performance results, respectively. Finally, Section VIIconcludes the paper with a brief discussion of future work.

II. RELATED WORK

There exists a considerable body of research available in theliterature that attempt to adapt to wireless channel fluctuations.Existing approaches can be classified into three categories:(1) rate adaptation, (2) MPDU length optimization, and (3)channel estimation.

Rate adaptation: Accurate rate selection in time-varyingwireless channel for WiFi systems has remained a challengingproblem. CARA [2] continuously tracks the success/failureof packet transmissions to select the most appropriate PHYrate that maximizes the throughput. SoftRate [3] and Accu-Rate [4] employ new methodologies which exploit the outputinformation of DSP blocks during the receiver processing inorder to enhance the accuracy and responsiveness against thechannel variation. However, these approaches rely on channelcharacterization at the granularity of frames (i.e., at the end ofeach frame reception), and hence, do not address the channeldynamics present within a frame. In [5], on the other hand,SIRA adapts the PHY rate on intra-frame basis to overcomethe channel dynamics. ChASER tracks the channel dynamics

while the receiver is actively processing the same frame, thusmaking it much more adaptive to changes that occur withinthe duration of a single A-MPDU frame.

MPDU length optimization: In [6], the authors attemptto optimize the A-MPDU subframe length according to thesignal-to-noise ratio (SNR) observed under the correspondingwireless channel condition. Their MPDU length selectionmechanisms are, however, effectively assuming that the SNR isuniform during the entire duration of the A-MPDU frame. Asa result, even a modest variation in channel dynamics fromthe effective SNR may lead to poor choice of the length.In [7], MoFA dynamically adapts the length of the A-MPDUby observing the increase of SFER in the latter part of A-MPDU. In contrast, ChASER is capable of keeping up withthe changes in channel behavior (under pedestrian mobility)regardless of the length of the A-MPDU frame, thus allowingfor the use of maximum length for maximum throughput.

Channel estimation: Numerous channel estimation algo-rithms have been proposed to enhance the accuracy of theestimation in a rapid dispersive fading channel environment.The work done in [8, 9] can be used to estimate the wirelesschannel based on the linear interpolation of partial channel in-formation, i.e., pilot signal. However, the technique is inappro-priate for 802.11n where the pilot subcarriers are not spacedclosely to estimate the channel fluctuation of data subcarriersin frequency domain. It is worth noting that the authors in [10]have proposed midamble-based channel estimation techniquesfor high mobility vehicular channel (e.g., 802.11p vehicle-to-vehicle communication specification [11]), and found toimprove the performance but suffer from two major disad-vantages: (1) protocol overhead incurred by addition of themidamble; and (2) deviation from the 802.11n standard. Inparticular, the latter makes it extremely costly and impracti-cal for large-scale adoption by commercial 802.11n vendors.Other notable techniques include spectral temporal averaging(STA) and constructed data pilot (CDP) estimation methodsintroduced in [12], both of which also use the estimatedtransmit data symbols which can be obtained based on theinformation from the receiver’s demapper to track the channelvariation. However, according to the results provided in [12],the error rate performance appears too high to be acceptable inpractical communication environment, because the informationis obtained in the absence of error correcting. The authorsin [13] have proposed the iterative channel estimation withpostamble, the concept of which is similar to that of theturbo channel estimation [14]. Unfortunately, the requirementof the iterative operation brings high computational complexity

0.32 ms

0.98 ms

1.64 ms

2.30 ms

2.96 ms

3.62 ms

4.27 ms

4.93 ms

5.59 ms

6.25 ms

6.91 ms

7.56 ms

8.22 ms

8.87 ms

9.53 ms

10.19 ms

0

1

0 0.1 0.2 0.3 0.4 0.5

CD

F

Amplitude change (∇A)

0

1

0 0.1 0.2 0.3 0.4 0.5

CD

F

Phase change (∇θ)

(a) Static trace

0

1

0 0.1 0.2 0.3 0.4

CD

F

Amplitude change (∇A)

0

1

0 0.1 0.2 0.3 0.4 0.5

CD

F

Phase change (∇θ)

(b) Mobile trace

Fig. 2: Temporal selectivity of the measured CSI with time gap, τ .

which is undesirable for WiFi. Moreover, the postamble causesincompatibility with the 802.11n standard.

In contrast, the channel estimation technique proposed byChASER is able to “chase” the channel dynamics within theA-MPDU duration up to its maximum length for pedestrianmobility, 802.11n standard-compliant, and of low computa-tional complexity with no additional overhead.

III. BACKGROUND AND OBSERVATIONS

In this section, we first describe briefly the IEEE 802.11nMAC/PHY features considered in this paper briefly. We thenquantify the changes of wireless channel behaviors affectedby pedestrian mobility based on the CSI measurement study.Finally, based on the extensive experiments, we show thesignificant degradation of WiFi performance when A-MPDUis enabled in the pedestrian mobile environment, and revealits causes which led to the design of ChASER.

A. Channel Estimation and Compensation

Due to fading, shadowing, and other interference sources,the transmitted signal through wireless channel experiencesmultiplicative distortions. To cope with them, many of theIEEE 802.11 standards (e.g., 802.11n and 802.11ac) requirethe receiver function to measure the CSI at the beginning ofthe reception using the PLCP preamble. As shown in Fig. 1,the 802.11n PLCP preamble includes legacy preamble (fromIEEE 802.11a) and high-throughput (HT) preamble (a newfield added in IEEE 802.11n).1

Each PLCP preamble carries short training field (STF),long training field (LTF), and signal (SIG) field. The legacySTF (L-STF) is used for signal detection, automatic gaincontrol (AGC), time synchronization, and coarse frequencyoffset estimation. L-LTF is composed of two repetitions ofknown long training sequences used for CSI estimation andfine frequency acquisition. On the other hand, the purposeof HT-STF is to improve AGC for MIMO system. HT-LTFalso aims to estimate CSI by considering the additional HTfeatures, i.e., MIMO and the increased number of subcarriers.In general, 802.11 devices commonly use least-squares (LS)estimation to obtain CSI, H , given by

H =YLTF

XLTF, (1)

1The 802.11n defines two PPDU formats, called mixed and greenfield. Thegreenfield format is rarely used due to the incompatibility with legacy devicesIn this paper, we consider the mixed format only.

where YLTF is the received LTF symbol corresponding to thetransmitted known LTF sequence, XLTF.

The receiver compensates the distorted OFDM data symbolsby using the CSI obtained from LTF. The extent that thecompensated received symbols in I-Q plane are dispersedfrom their ideal positions depends on the accuracy of theestimated CSI. Therefore, if wireless channel varies swiftlyand considerably during the reception of a single frame, theCSI obtained at the PLCP preamble is likely to be inadequateto compensate the channel for the symbols belonging to thelatter part of the frame.

B. Role of Pilot Subcarriers

Signal distortion between transmitter and receiver is causednot only by wireless channel, but also by symbol timingoffset (STO). The STO stems from clock mismatch betweentransmitter and receiver, and is linearly proportional to thesubcarrier index [15]. In 802.11n 20 MHz OFDM symbol,as shown in Fig. 1, four evenly spread pilot subcarriers areinserted to transmit predefined information [16]. The phasedifference caused by STO can be eliminated by interpolatingthe measured phase offset values from these pilot symbols.

However, pilots are insufficient to track the CSIs for allthe subcarriers. Typically, the maximum root mean square(RMS) channel delay spread (στ ) considered in 802.11 WLANis about 250 ns long, corresponding to the 90% coherencebandwidths,2 Bc,90% = 1/50στ = 80 kHz, which is muchsmaller than the spacing between two consecutive pilots infrequency domain, i.e., 4,375 kHz [17]. Thus, the receivercannot estimate the CSI of all the subcarriers using pilotsubcarriers alone.

In summary, even though there are mechanisms designed totrack the channel status for the entire frame, i.e., preamble andpilots, if the channel varies significantly during the durationof the frame, the CSI obtained from them might not besufficiently accurate to decode a long frame, especially whenthe transmission duration is much longer than the channelcoherence time.

C. Frame Aggregation

To improve the network efficiency, 802.11n and 802.11achave defined frame aggregation scheme, e.g., A-MPDU [16,18]. When applying A-MPDU, the transmission is conducted

2Bc,x% is the bandwidth where the autocorrelation of the channel in thefrequency domain is equal to x% of the peak.

0

0.2

0.4

0.6

0.8

1

1 2 3 4 5 6 7 8

SF

ER

Subframe location (ms)

IWL5300 (1 m/s)IWL5300 (0.5 m/s)IWL5300 (0 m/s)AR9380 (1 m/s)AR9380 (0.5 m/s)AR9380 (0 m/s)

Fig. 3: Measured robustness of A-MPDU. The subframelocation indicates the beginning of the transmission of thecorresponding subframe relative to the beginning of the entireframe in ms.

on the basis of a frame consisting of several aggregatedMPDUs, by which the proportion of the protocol overheads,such as channel access delay, is reduced compared to the singleMPDU-based legacy protocol. Each MPDU is called subframeof the A-MPDU. Moreover, A-MPDU allows each subframeto be individually acknowledged and selectively retransmittedthanks to the usage of cyclic redundancy check (CRC) at theend of each subframe.

However, the aggregation leads to longer frame duration andless correlation between CSI of the first and the last parts ofthe frame. Specifically, the default maximum PPDU duration,Tmax, of 802.11n (10 ms) is likely to be longer than thechannel coherence time.

D. Temporal Selectivity

We now analyze the temporal selectivity of the real wirelesschannel to quantify the degree of variation of wireless channelduring Tmax, i.e., 10 ms in static and mobile scenarios.

We first use Intel WiFi Link (iwl5300) network interfacecard (NIC) and the modified device driver to collect CSI tracefrom HT-LTF of each successfully received frame [19, 20].The detail of the experimentation setting is described inSection VI-A To build the actual channel response experiencedby data OFDM symbols, we eliminate the phase offset inducedby STO or sampling clock difference between transmitter andreceiver, which is the same functionality of pilot subcarriersas described in the previous section. For the mobile scenario,we walked with the device at approximately 1 m/s speed.

We employ the following metrics which represent the nor-malized amplitude changes (∇A) and phase changes (∇θ) toevaluate the temporal selectivity [21]:

∇A(τ) =‖A(t)−A(t+ τ)‖2‖A(t+ τ)‖2

∇θ(τ) =‖θ(t)− θ(t+ τ)‖2√∑

Nscπ2

,

(2)where the vectors A(t) and θ(t) represent all the subcarriers’amplitudes and phases of the frame received at t, respectively,and τ is the time gap. ‖·‖2 is the l2-norm, and τ varies from250 µs to 10 ms. In this case, phase change is normalized byπ, of which 10% change is equivalent to 0.314 radians or 18◦.

Fig. 2 shows the cumulative distribution functions (CDF) of∇A(τ) and∇θ(τ) between the CSI of two frames with varying

AP C1 C2

C3

E1

C4C5

C6

CSI collection (Sec. V-A) Experiment (Sec. III-E)

0 2 4 6 m

Fig. 4: Topology.

τ . As shown in Fig. 2(a), ∇A(τ) and ∇θ(τ) remain relativelysteady in the static scenario. More than 80% of samples showthe changes within 10% even if τ is 10 ms.

However, the amplitude and phase variations increase withτ in the mobile scenario as shown in Fig. 2(b). When τ is setto 10 ms, ∇A(τ) and ∇θ(τ) vary by more than 10% betweentwo CSIs for over 99% and 95% of samples, respectively.

Consequently, both amplitude and phase response of thelatter part frame are expected to considerably deviate from thevalues predicted by the CSI estimated at the PLCP preamblewhen using A-MPDU, and hence the symbols at the latterpart is compensated by a false CSI. Thus, the preamble-basedchannel estimation is highly unreliable when employing A-MPDU and users are mobile.

E. Unreliability of A-MPDU in Mobile Environments

In order to verify the observation that the latter part of anA-MPDU is likely to suffer from more severe decoding error,we have conducted measurement using off-the-shelf 802.11ndevices, which support A-MPDU. We use the AP describedin Section V-A and use both Qualcomm Atheros ar9380 andiwl5300 NICs on the station side. Our experiments operate at5.22 GHz unlicensed band in a controlled office environments;no interference is observed.

We place the station device approximately 5 m away fromthe AP (E1 in Fig. 4), where the distance is relatively shortsuch that frames can be transmitted with the highest rate (i.e.,MCS 7 with 64QAM and 5/6 code rate) almost perfectly. Then,we carry the station and walk towards the AP at two pedestrianspeeds of 0.5 and 1 m/s until we arrive at the AP. We let theAP transmit frames whose size is 1,538 bytes continuouslywith MCS 7. The results are averaged over 5 runs.

Fig. 3 shows the SFER performance for three differentmobility scenarios (0, 0.5, and 1 m/s) for two different WiFidevices. SFER represents the frame error rate of a certainsubframe, through which we can examine the relationshipbetween subframe error rate and the subframe location. In thestatic case (0 m/s), the SFER holds relatively steady regardlessof the subframe locations. However, when we walk towardsthe AP, even though the AP and station are getting closer, theSFER dramatically increases along with the increase in the gapbetween the preamble and the respective subframe. Moreover,the slope of SFER curves becomes higher as the speed of the

0

0.1

0.2

0.3

0 500 1000 1500 2000

EV

M

OFDM Symbol Index

0 m/s0.3 m/s0.6 m/s1 m/s

(a) Pilot symbol dispersion.

I

Q

(b) Data symbol dispersion atthe front part of A-MPDU.

I

Q

(c) Data symbol dispersion atthe latter part of A-MPDU.

Fig. 5: Symbol dispersion increases in the presence of usermobility. All symbols are corrected by CSI and pilots.

station increases. We call such losses caudal losses as theytend to appear more frequently towards the end of the frame.

F. Relation between Symbol Dispersion and Mobility

In order to better understand the root causes of caudallosses, we investigate how badly the OFDM symbols aredistorted by the use of CSI from the preamble in mobilescenarios. This experiment have been conducted using theSora platform which contains the open source implementationof 802.11a specification . The details of implementation andexperimental setting are described in Section V-A. Let Ybe the compensated received symbol position and X be itsideal position in I-Q plane. The amount of symbol dispersionis then defined as |Y−X||X| , so-called error vector magnitude(EVM) [23].

Fig. 5(a) shows the average pilot symbol dispersion for eachOFDM symbol when the sender moves at an approximatelyconstant speed of 0, 0.3, 0.6, and 1 m/s. Note that since wemust know the symbol’s ideal position, X , to calculate EVM,we use the known pilot symbols instead of unknown datasymbols. In mobile environments, the dispersion increases asOFDM symbol index increases, which means the compensatedsymbol moves gradually farther away from its ideal position.Furthermore, the slope of EVM curves becomes higher as theaverage speed of the sender increases. However, when thesender holds its location, the dispersion is stable regardless ofOFDM symbol index. In the case, since the dispersion is onlyaffected by the random noise, the wireless channel variationbetween two transceivers is quite stable.

Figs. 5(b) and 5(c) show the received data symbol positions

Sniff

Preamble

insertion

Pilot

insertion

Mapper

Interleaver

Encoder

Scrambler

MAC layer (CRC check)

S/P, FFT, RF front-end

Decoder

De-

interleaver

Demapper

Pilot-based

equalizer

CSI-based

equalizer

Descrambler

Channel

Estimator

Adaptive

Filter

Tx S

ym

bo

l R

eco

nstr

ucto

r

CR

C

Qu

eu

e

Queue

X

X

Y

H

DD

D

Fig. 6: Block diagram of ChASER. The dashed block andlines are the additional processing needed for ChASER.

derotated by CSI from LTF when the sender moves at anaverage speed of 1 m/s. In the case of front part symbols, thoseare located in the vicinity of 16QAM signal constellations. Onthe other hands, the symbol dispersion at the latter part of A-MPDU is much larger than that at the front part of A-MPDU.

As we explain in Section III-A, the dispersion mainlydepends on the accuracy of the CSI used for compensatingthe distorted symbols. Therefore, we conclude that the caudallosses are induced by inaccurate CSI which is only estimatedat the beginning of A-MPDU and pilots are insufficient to trackthe channel variation during a specific A-MPDU duration.

IV. CHASER DESIGN

In this section, we describe the ChASER scheme designedto minimize caudal losses by tracking the channel accuratelyduring the course of demodulation/decoding while satisfyingthe latency requirement, i.e., completing the frame receptionprocess within the SIFS (16 µs) interval.

ChASER is composed of three main blocks: channel esti-mator using unknown data symbols, adaptive filter, and CRC-assisted channel corrector.

A. Channel Estimation using Unknown Data Symbols

As shown in Fig. 6, the received signal (Y ) is processedby demapper and decoder sequentially, whose output are Dand D, respectively. In order to exploit the error correctinggain of the channel coding, we reconstruct the original OFDMsignal based on the D rather than D , based on which STAand CDP conduct the reconstruction without error correct-ing gain [12, 24]. ChASER re-encodes and re-modulatesD as done by the DSP block of the transmitter, which iscomposed of encoder, interleaver, and mapper. Note that theknown sequence is inserted for pilot subcarrier. We refer tothe block as symbol reconstructor. Afterwards, utilizing thesame methodology as the preamble-aided channel estimationscheme as described in Section III-A, we can estimate CSIby replacing the denominator of LS estimator (i.e., Eq. (1))

from known sequence (XLTF) with the reconstructed OFDMsymbol (X) as follows:

Hi =Yi

Xi

, (3)

where Hi and Yi represent the estimated CSI and the receivedsymbol of the ith OFDM symbol, respectively.

Note that by using the transmit processing blocks of 802.11nas shown in Fig. 6, no additional processing blocks are neededin ChASER design. Moreover, thanks to the error correctingcapability of the channel decoder, ChASER estimates the CSIwith high fidelity.

B. Adaptive Filter

Unfortunately, even if the original transmit symbols arereconstructed with the help of the channel coding, the esti-mation errors might still exist, which will in turn result inerror propagation and performance degradation.

To mitigate the error propagation, we develop an adaptivefilter using exponential weighted moving average (EWMA),with step size equal to µ:

Hi = (1− µ)Hi−1 + µYi

Xi

. (4)

The choice of µ has a trade-off between the estimation errorand the convergence speed.

Proposed filter can smoothen out the sharp edge of Hi

which comes from estimation error, and hence we can alleviateerror propagation and the resultant performance degradation.Furthermore, one of the merit of the proposed filter is thatit only requires memory storage for the CSI estimated at thepreceding OFDM symbol, Hi−1, and has low computationalcomplexity.

Algorithm 1: CSI update of ChASER.

input : XLTF, Xi, Yi, Xi−1, Yi−1, Hi−1

output: Hifor each OFDM symbols i do

if is HT-LTF then/* i = 0 */Hi ← Yi

XLTF

else if includes CRC and passes CRC check thenif has data bits of the next subframe then

/* Xi−1 is known */Hi−1 ← Yi−1

Xi−1

Hi ← (1− µ)Hi−1 + µ Yi

Xi

else/* Xi is known */Hi ← Yi

Xi

endelse

/* unknown data symbol */Hi ← (1− µ)Hi−1 + µ Yi

Xi

endend

0

0.04

0.08

0.12

0.01 0.05 0.1 0.2 0.3 0.4 0.5

EV

M

µ

Fig. 7: EVM performance with various µ values.

C. CRC-assisted Channel Correction

It is generally accepted that A-MPDU is very efficient inerror-prone environments thanks to the selective retransmis-sion capability. To achieve this, A-MPDU attaches CRC at theend of each subframe to detect bit error as shown in Fig. 1. Letus consider the case when a subframe is successfully received;no bit error observed. Since all the bits, D, of the subframe aredecoded correctly, the reconstructed symbols are exactly thesame as the original transmitted symbols. Therefore, we canobtain the error-free CSIs from unknown data OFDM symbols,which are known symbols from now on. In this case, ChASERupdates the estimated CSI, Hi, based on the LS estimator,Eq. (3), instead of the proposed adaptive filter, Eq. (4).

This achieves the same effect that the transmitter oppor-tunistically injects LTF at the end of every subframe, calledmidamble [10]. However, our approach has no extra protocoloverhead, which wastes air time.

Algorithm 1 summarizes the overall CSI update procedureof ChASER depending on whether current symbol is knownor not. First, if the current symbol is known, such as HT-LTF or all data bits decode correctly, the LS estimator isused. If not, the adaptive filter is used. However, even thoughthe symbol contains CRC and the corresponding subframepasses CRC check, we cannot be assured that all data bits aredecoded correctly if CRC is followed by data bits of the nextsubframe, which does not pass CRC check yet. In such case,there is potential error in the current OFDM symbol, whereasthe preceding OFDM symbol should be known, because allthe containing bits pass CRC check. Hence, the CSI of theknown preceding symbol, Hi−1, is first updated based on theLS estimator, and then the current CSI, Hi, is estimated byusing the adaptive filter.

D. Impact of Step Size µ

Before verifying the benefits of ChASER, we investigatethe impact of step size, µ, which directly affects the perfor-mance of ChASER. For larger values of step size, the trackingspeed is very fast, but the steady state estimation error islarge and vice versa. To determine an appropriate µ, we runChASER with various values on all collected traces of thetrace-driven link level simulator. We observed the best EVMperformance with µ = 0.1 as shown in Fig. 7, and hence weset µ as 0.1 throughout the paper.

V. TESTBED EXPERIMENTS

A. Implementation and Testbed Settings

To verify the feasibility of the proposed scheme, we haveimplemented ChASER using the Sora SDK version 1.6,

0

0.2

0.4

0.6

0.8

1

0 10 20

SF

ER

Subframe Index

Baseline 802.11n - 0 m/s

Baseline 802.11n - 0.3 m/s

Baseline 802.11n - 1 m/s

ChASER - 0 m/s

ChASER - 0.3 m/s

ChASER - 1 m/s

(a) Subframe error ratio.

0

10

20

30

40

Baseline 802.11n ChASER

Th

rou

gh

pu

t (M

bit

/s) 0 m/s 0.3 m/s 0.6 m/s 1 m/s

(b) Throughput.

Fig. 8: Performance comparison in Sora.

which supports 802.11a only. We created and inserted severalessential features of the IEEE 802.11n standard such as HT-preamble, increased number of subcarriers, and A-MPDU,and embedded them as part of the platform. The operationof ChASER is pipelined with DSP functions of the IEEE802.11n over multiple cores based on Sora User-Mode Exten-sion (UMX) as specified in Section IV to satisfy the latencyconstraint. Due to some limited capabilities of the decoderof Sora SDK, ChASER updates the CSI at every 4 OFDMsymbols in our implementation.

Our testbed operates at 5.22 GHz band and all the exper-iments have been done under a clean channel environment;no interference signal was detected throughout the test. Wefirst install the programmable AP equipped with ar9380 NIC,ath9k, and hostap [25–27]. We let Sora operate as the receiverof the AP, whose downlink buffer is always filled with A-MPDUs, each of which consists of several 1000-byte MPDUs.Since Sora testbed is too heavy to move swiftly, we move theAP at an approximately pedestrian speed on average, e.g., 0,0.3, 0.6, and 1 m/s, and sets the transmission rate as MCS 3(16QAM and 1/2 code rate).3 Basically, Tmax is set to 10 msthroughout the paper. The following results are averaged over5 runs, where each run lasts for about 100 seconds.

B. Performance Comparison

Figs. 8(a) and 8(b) plot the results of experiments in termsof SFER and throughput, respectively. From the results, wecan argue that the baseline 802.11n is quite vulnerable to the

3Since Sora is inherently lack of fine AGC operation, we tuned the receivergain offline to an appropriate value. Thus, the AP should move over relativelyshort distances (e.g., 1 m) so that the tuned value can be valid for a while.Moreover, because the capabilities of RF front-end of Sora is far from beingdecent, it is difficult to use a higher order MCS where ChASER may achievesignificant performance gain over the baseline 802.11n.

0.5

0.01

0.1

0 500 1000 1500 2000

Esti

mati

on

err

or

OFDM Symbol Index

Baseline 802.11nMidamble (γ = 10)Midamble (γ = 200)

STAChASER

(a) Estimation accuracy with fixed MCS 7. 0.5

0.01

0.1

0 500 1000 1500 2000

Esti

mati

on

err

or

OFDM Symbol Index

Baseline 802.11nMidamble (γ = 10)Midamble (γ = 200)STAChASER

(b) Estimation accuracy with fixed MCS 1.

Fig. 9: Channel estimation accuracy in mobile traces.

unexpected side effect of user mobility while ChASER caneffectively enhances the throughput performance and reducethe SFER by conducting more accurate channel compensation.That is, ChASER chases the wireless channel variation withhigh fidelity. Consequently, the maximum 58% throughputgain can be achieved by adopting ChASER.

VI. SIMULATION

In order to demonstrate the performance gain of the pro-posed scheme more extensively, we have conducted trace-driven link level simulations.

A. Simulation Methodology

All features of IEEE 802.11n PHY/MAC are embeddedin our trace-driven link level simulator by using GNU Ra-dio and IT++ libraries [28, 29]. We collect the fine-grainedchannel traces using iwl5300 NIC and the 802.11n CSI toolon laptop [20] which reports the measured CSI for everysuccessfully received packet [19, 20]. We collect 14 traceswith various speeds of the receiver at different locations (redpoints in Fig. 4). Each CSI contains the complex channelfrequency responses, which represent amplitude and phasedifferences between transmitter and receiver, of 30 OFDMsubcarriers.4 The channel model of the simulator is builtby linearly interpolating the measured CSIs over time andfrequency domains.

Moreover, midamble and STA compared as the existingschemes have been also implemented and embedded into thesimulator. In midamble, the training OFDM symbols modu-lated with BPSK are inserted periodically with the interval

4The device driver we used just reports the channel responses of a groupof 30 subcarriers instead of all the 56 subcarriers.

0

20

40

60

80

0 2 4 6 8 10

Th

rou

gh

pu

t (M

bit

/s)

Maximum PPDU Duration (ms)

Baseline 802.11nMidamble (γ = 10)Midamble (γ = 200)ChASERUpper bound

Fig. 10: Impact of Tmax with fixed MCS 7.

of γ OFDM symbols. Note that midamble can achieve betterestimation accuracy if the training symbols are inserted morefrequently. That is, there is a trade-off between the estimationaccuracy and the protocol overhead generated by the trainingsymbols in midamble. We select two training symbol intervals,i.e., 10 and 200 OFDM symbols, as two representatives of amore accurate one and a more efficient one, respectively.

B. Simulation Results

Estimation accuracy: Fig. 9 shows the channel estimationaccuracy of various schemes in terms of EVM using thechannel trace collected at C1 in Fig. 4, a relatively goodchannel environment compared to other spots. Fig. 9(a) andFig. 9(b) represent the results when the transmission rates arefixed as MCS 7 and MCS 1 in a log-scale, respectively.

As expected, the EVM of baseline 802.11n continuouslyincreases along with the increase of the OFDM symbol index,thus resulting in caudal losses ultimately. In the case ofMidamble200,5 the EVM value moves upward and downwardperiodically, reflecting the estimation error increase as timeelapses and the periodic channel updates done by trainingsymbols, respectively.

ChASER achieves similar performance to Midamble10which shows the best estimation performance as shown inFig. 9, while the latter incurs much more protocol overheadscompared to the former. In the case of STA, because it uses theoutput of the demapper rather than that of the channel decoder,channel coding gain is not exploited and the performanceis highly vulnerable to the demodulation error. As a result,STA can achieve better performance in the case of MCS 1(Fig. 9(b)) than in the case of MCS 7 (Fig. 9(a)).

TABLE I: EVM performance in various SNR channel traces.Low SNR High SNR

Modulation QPSK 16QAM 64QAM QPSK 16QAM 64QAM

Midamble10 0.0367 0.0371 0.0370 0.0210 0.0207 0.0206Midamble200 0.0612 0.0617 0.0614 0.0380 0.0379 0.0378

STA 0.1098 0.4881 0.5194 0.0848 0.2265 0.3721ChASER 0.0398 0.0459 0.0499 0.0248 0.0250 0.0292

Table I presents the average EVM values when applying var-ious SNR channel traces. Midamble10 shows the lowest EVMby introducing tremendous training overhead, while ChASERshows similar EVM values (at most 1.3% difference) withoutadditional training symbols.

5MidambleX is the midamble whose interval γ is equal to X .

0

1

0 10 20 30 40 50 60

CD

F

Throughput (Mbit/s)

Baseline 802.11nMidamble (γ = 10)Midamble (γ = 200)ChASER

(a) Static traces.

0

1

0 10 20 30 40 50 60

CD

F

Throughput (Mbit/s)

Baseline 802.11nMidamble (γ = 10)Midamble (γ = 200)ChASER

(b) Mobile traces.

Fig. 11: Throughput distribution for all the collected traces.

Impact of the A-MPDU duration: We then investigate howthe A-MPDU length will influence the throughput when we ap-ply various channel estimation schemes described earlier. Weuse MCS 7 and vary the duration Tmax of a single A-MPDUtransmission. Fig. 10 shows the throughput performance.6

When Tmax is less than 2 ms, all the schemes includingthe baseline 802.11n yield almost the same throughput per-formance approaching the upper bound,7 because the wirelesschannel is quasi-stationary within the transmission time. WhenTmax exceeds 2 ms, the performance gaps between the upperbound and the existing schemes become larger while ChASERstill achieves a performance close to the upper bound. Partic-ularly, ChASER improves throughput up to 99.5% comparedto the baseline 802.11n in this result. It is worth noting thatMidamble10, which shows the best estimation accuracy inFig. 9, has less throughput gain than the proposed schemedue mainly to the protocol overhead incurred by midamble.Throughput performance: In fact, ChASER is an inde-pendent block which can be combined with a rate controlalgorithm so that both frame-level and symbol-level resilienceto channel error are enhanced together. We have embeddedRRAA, one of the best rate control algorithms, into thesimulator on top of various channel tracking algorithms andextensively evaluated the throughput over all the collectedtraces. Fig. 11 plots the CDF of the throughput with variouschannel tracking algorithms. The baseline still shows the worstperformance for all the cases even if RRAA chooses the ap-propriate PHY rate. We observe that the lines corresponding toChASER are always right most for relatively high throughputregion, e.g., from 50 Mbit/s to 60 Mbit/s, which is likelyto be the region most applications are of interest. In otherwords, a user can be served with better throughput with higherprobability when ChASER is used.

At last, in order to evaluate throughput performance in amore general environment, we conduct network-level simula-tion, letting six nodes be located at six CSI collection pointswith various levels of mobility and associate with the AP asshown in Fig. 4. We make them share the wireless mediumupon carrier sense multiple access with collision avoidance(CSMA/CA) and be able to sense each other’s signal so that

6In this simulation, we did not consider STA because it shows undesirablethroughput performance (almost zero with MCS 7)

7The upper bound is achieved when there is no transmission error.

0

5

10

15

0 m/s 0.5 m/sHigh SNR

1 m/s 0 m/s 0.5 m/sLow SNR

1 m/s

Th

rou

gh

pu

t (M

bit

/s)

Baseline 802.11nMidamble (γ = 10)

Midamble (γ = 200)ChASER

Fig. 12: Multi-node environments.

there is no hidden terminal problem during the simulation.Fig. 12 presents the simulation results when six nodes

always have uplink A-MPDU to transmit. We divide the sixnodes into two groups indicated here as high SNR and lowSNR based on their average SNR values and intentionallycontrol the mobility level of each node so that the three nodesof each group have different mobility levels. We observe thatChASER and Midamble10 outperform all the other schemes inall the cases and the former is better than the latter except thecase of the static node with low SNR. The reason we find out isthat the step size of ChASER, µ, is not selected appropriatelyin this case, which is one of the issues that we need to addressin the future work. In short, ChASER achieves 79.8% and23.5% higher throughput compared to baseline 802.11n formobile and static nodes, respectively.

VII. CONCLUSION AND FUTURE WORK

This paper presents ChASER, a low-complexity, standard-compliant channel estimation and tracking design that achievesremarkably high performance even in the presence of pedes-trian user mobility and frame aggregation. The primary contri-bution of ChASER is to reconstruct transmit symbols at thereceiver with extremely low estimation errors by exploitingthe output of decoder and the existing transmit DSP blockswith low-complexity adaptive filter. The reconstructed datasymbols are then used to estimate and track the CSI at thegranularity of data symbols, leading to significantly improvedrobustness against the time-varying channel dynamics over theduration of the aggregated frame. We demonstrate its benefitsand feasibility via software-radio prototype implementationand trace-driven link level simulator which reflect the wirelesschannel characteristics of the real world. We envision theChASER’s channel tracking scheme to be applicable to anycommunication system that employs coherent modulation, notjust WiFi devices considered in this paper.

We conclude the paper with a brief discussion of ongoingresearch on further improving and expanding ChASER.Adaptive µ: The larger µ is, the faster ChASER operatesbut at the increasing risk of higher estimation errors for staticscenarios. To optimize ChASER for both performance andspeed, it is desired to adapt µ according to the mobility pattern.We plan to address this issue in our future work.MIMO extension: We plan to extend ChASER’s CSI estima-tion scheme to MIMO-enabled WiFi devices. Since the impactof CSI estimation accuracy is more pronounced for MIMO, we

envision that the expected performance gain by this extensionmay lead to even more significant performance improvement.

ACKNOWLEDGMENT

This paper was partially studied with the support of theMinistry of Science, Ict & future Planning (MSIP), andpartially supported by the Brain Korea 21 Plus Project in 2014.

REFERENCES[1] W. Sun, O. Lee et al., “Wi-Fi Could Be Much More,” IEEE Commun.

Mag., vol. 52, no. 11, pp. 22–29, 2014.[2] J. Kim, S. Kim et al., “CARA: Collision-Aware Rate Adaptation for

IEEE 802.11 WLANs,” in Proc. IEEE INFOCOM, 2006.[3] M. Vutukuru, H. Balakrishnan, and K. Jamieson, “Cross-layer Wireless

Bit Rate Adaptation,” in Proc. ACM SIGCOMM, New York, New York,USA, 2009.

[4] S. Sen and N. Santhapuri, “AccuRate: Constellation based Rate Estima-tion in Wireless Networks,” in Proc. USENIX NSDI, 2010.

[5] O. Lee, J. Kim, and S. Choi, “SIRA: SNR-aware Intra-frame RateAdaptation,” accepted to IEEE Commnu. Lett., 2015.

[6] K. Feng, P. Lin, and W. Liu, “Frame-Aggregated Link AdaptationProtocol for Next Generation Wireless Local Area Networks,” EURASIPJournal on Wireless Communications and Networking, pp. 1–13, 2010.

[7] S. Byeon, K. Yoon et al., “MoFA: Mobility-aware Frame Aggregationin Wi-Fi,” in Proc. ACM CoNEXT, Dec. 2014.

[8] A. Stamoulis, “Intercarrier Interference in MIMO OFDM,” IEEE Trans.Signal Processing, vol. 50, no. 10, pp. 2451–2464, 2002.

[9] C. Y. Shin and E. J. Powers, “Double Selective Channel Estimation forOFDM Systems,” in Proc. IEEE Asilomar, Nov. 2005.

[10] S. I. Kim, H. S. Oh, and H. K. Choi, “Mid-amble Aided OFDMPerformance Analysis in High Mobility Vehicular Channel,” in Proc.IEEE Intelligent Vehicles Symposium, Jun. 2008.

[11] IEEE 802.11p, Part 11: Wireless LAN Medium Access Control (MAC)and Physical Layer (PHY) specifications: Wireless Access in VehicularEnvironments, IEEE Std., Jul. 2010.

[12] Z. Zhao, X. Cheng et al., “Channel Estimation Schemes for IEEE802.11p Standard,” IEEE Intell. Transport. Syst., vol. 5, no. 4, pp. 38–49,Oct. 2013.

[13] T. Zemen and L. Bernado, “Iterative Time-Variant Channel Estimationfor 802.11p Using Generalized Discrete Prolate Spheroidal Sequences,”IEEE Trans. Veh. Technol., vol. 61, no. 3, pp. 1222–1233, 2012.

[14] M. Zhao, Z. Shi, and M. C. Reed, “Iterative Turbo Channel Estimationfor OFDM System over Rapid Dispersive Fading Channel,” in Proc.IEEE ICC 2007, Jun. 2007.

[15] E. Perahia and R. Stacey, Next Generation Wireless LANs: Throughput,Robustness, and Reliability in 802.11n. Cambridge Univ. Press, 2008.

[16] IEEE 802.11, Part 11: Wireless LAN Medium Access Control (MAC)and Physical Layer (PHY) specifications, IEEE Std., Mar. 2012.

[17] IEEE P802.11 Wireless LANs, TGn Channel Models, IEEE 802.11-03/940r4, May 2004.

[18] IEEE 802.11ac, Part 11: Wireless LAN Medium Access Control (MAC)and Physical Layer (PHY) specifications: Enhancements for Very HighThroughput for Operation in Bands below 6 GHz, IEEE Std., Dec. 2013.

[19] Intel Wi-Fi Link 5300. http://www.intel.com/products/wireless/.[20] Linux 802.11n CSI Tool. http://dhalperi.github.io/linux-80211n-csitool/.[21] A. Bhartia, Y. C. Chen, and L. Qiu, “Harnessing Frequency Diversity

in Wi-Fi Networks,” in Proc. ACM MobiCom’11, Sep. 2011.[22] K. Tan, J. Zhang et al., “Sora: High Performance Software Radio Using

General Purpose Multi-core Processors,” in Proc. USENIX NSDI, 2009.[23] A. Georgiadis, “Gain, Phase Imbalance, and Phase Noise Effects on

Error Vector Magnitude,” IEEE Trans. Veh. Technol., vol. 53, no. 2, pp.443–449, March 2004.

[24] A. Bourdoux, H. Cappelle, and A. Dejonghe, “Channel Tracking forFast Time-Varying Channels in IEEE 802.11p Systems,” Proc. IEEEGlOBECOM, Dec. 2011.

[25] AR9380. http://www.qca.qualcomm.com/mobile-connectivity/wi-fi/.[26] Ath9k: Atheros Linux Wireless Driver. http://wireless.kernel.org/en/

users/Drivers/ath9k/.[27] HostAP: IEEE 802.11 AP, IEEE 802.1X/WPA/WPA2/EAP/RADIUS

Authenticator. http://hostap.epitest.fi/hostapd/.[28] GNU Radio. http://http://gnuradio.org/.[29] IT++. http://itpp.sourceforge.net/.