JACPoT: A Joint Algorithm for Channel and PowerAssignment in Enterprise Wireless Networks

Seongwon Kim, Yeonchul Shin, Seungmin Yoo, and Sunghyun ChoiDepartment of ECE and INMC, Seoul National University, Seoul, Korea

{skim11, ycshin, smyoo}@mwnl.snu.ac.kr, schoi@snu.ac.kr

Abstract—The demand for wireless communication in enter-prise network has been increasing with the advent of “smartworking environment.” In enterprise wireless network, highdensity of access points (APs) and limited frequency resourcesraise the significance of radio resource management. In thispaper, we propose a centralized algorithm, named JACPoT, forjointly executable transmission power control (TPC) and dynamicchannel assignment (DCA). The proposed algorithm first guar-antees that the links between AP and client stations are robust tothe inter-cell interference via TPC. Next, it minimizes contentionbetween neighboring APs via DCA utilizing a novel metric andannealed Gibbs sampler. Through ns-3 simulations based onmeasurement in real office environments, we demonstrate thatJACPoT reduces uplink frame collisions by 36% and significantlyenhances both data and voice traffic delivery.

Index Terms—Enterprise network, IEEE 802.11 WLAN, dy-namic channel assignment, transmit power control.

I. INTRODUCTION

The skyrocketing demand for high performance wirelessaccess has raised the density of wireless LAN (WLAN)access points (APs). With the existence of many APs andnon-AP stations (STAs), the significance of radio resourcemanagement (RRM) is more emphasized. Accordingly, IEEE802.11ax task group has been recently launched to improvespectral efficiency of WLAN, especially, in dense deploymentscenarios [1].

At the same time, the demand for enterprise WLAN sup-porting various office applications has been increased withthe advent of new working environments using diverse smartdevices. Since enterprise WLAN requires stable performancein delivering both heavy traffic and real-time applications, APsare deployed very densely in general. The communication insuch densely deployed WLAN, however, suffers from severeinter-cell interference (ICI). To be more specific, the ICIlouder than clear channel assessment (CCA) threshold, whichwe call exposed interference in this paper, degrades networkperformance by intensifying contention. On the other hand, theICI quieter than CCA threshold, called hidden interference, de-grades the performance by deteriorating signal-to-interference-plus-noise-ratio (SINR).

In the above-mentioned enterprise environment, the radioresource can be more advertently managed thanks to theexistence of AP controller (APC). As illustrated in Fig. 1,APs are connected to an APC, and send feedback information,e.g., received signal strength (RSS) and traffic load, to theAPC. Then the APC conducts RRM operation such as transmit

Fig. 1. An architecture of enterprise WLAN.

power control (TPC) and dynamic channel assignment (DCA)in a centralized manner. Specifically, TPC schemes assignappropriate transmit powers to each AP to adapt environmentalchanges in network. DCA schemes, on the other hand, dy-namically assign frequency channels to each AP to maximizespatial reuse in an adaptive manner.

The goal of our study is to propose a jointly executableTPC and DCA algorithm specialized in enterprise wirelessnetwork. We first propose a simple TPC scheme that guaran-tees successful packet transmissions by assigning appropriatetransmit power to endure hidden interferences and minimizeICI. Secondly, we propose a novel metric for network perfor-mance which precisely reflects performance degradation dueto the contention among co-channel APs, i.e., the APs operatein the same frequency channel. Combining the proposedTPC scheme and metric, and then adopting a meta-heuristicalgorithm, we finally propose JACPoT (Joint Algorithm forChannel and Power assignmenT).

In contrast to the existing work, JACPoT (i) jointly assignsfrequency channels and transmit powers, and (ii) considersthe physical asymmetry between AP and STA devices such asdifferent receive sensitivity, antenna gain, etc. Through ns-3simulation based on the measurement results in real officeenvironments, the performance of JACPoT is comparativelyevaluated with existing schemes.

The rest of the paper is organized as follows. In Section II,we review the existing schemes for TPC and DCA. Section IIIintroduces JACPoT in a nutshell. Section IV and Section Vdescribe the proposed TPC and DCA schemes, respectively.Simulation-based performance evaluation is presented in Sec-tion VI, and Section VII concludes the paper.

II. RELATED WORK

We first introduce existing TPC and DCA schemes. Sincewe consider enterprise WLAN where APC manages connectedAPs, we focus more on centralized schemes.

A. TPC Schemes

There have been remarkable studies on TPC algorithms inthe literature. Power control for AP performance enhancement(PCAP) proposed in [2] consists of two steps. The first (sec-ond) step aims to maximize (maximize) the mean (variance)of AP utility, defined as the weighted product of its associatedSTAs’ datarates. These datarates are predicted by calculatingthe worst case SINR and using SINR-datarate table.

In [3], S. Bae et al. propose an RRM scheme including TPCand DCA. The proposed TPC scheme controls the transmitpower so that the RSS of STAs in neighboring cells becomesunder the CCA threshold to reduce unnecessary contentions.However, the scheme does not consider the channel variationcaused by multi-path fading.

In Cisco’s TPC version 1 [4], AP reduces transmit powerif RSS of its third loudest neighbor AP is larger than a TPCthreshold. This scheme does not consider different channelsof APs, and the performance highly depends on the selectionof TPC threshold, which is configured manually.

B. DCA Schemes

In this section, we introduce several DCA schemes in theliterature. In [5], the proposed scheme aims at minimizingthe signal strength from neighboring APs at a reference AP.With given channels, neighboring APs take turns to selectthe operating channel which yields the minimum interferenceto the reference AP. The authors of [6] propose a heuristicalgorithm to minimize an objective function, the effectivechannel utilization of the most stressed bottleneck AP, basedon interferer classification. This heuristic algorithm repeatedlyassigns the best channel to the most stressed AP until no betterchannel plan exists.

In [7], the authors propose a novel metric for DCA by aproduct of RSS, interference factor, and channel occupancytime (COT). A distributed algorithm is then proposed tominimize the metric for each AP by changing channel set. InCisco’s DCA scheme [4], a cost metric (CM) for an AP is firstdefined as a function of RSS, noise, interference, utilization,and traffic load. According to the CM, the worst AP in an RFsubgroup and its first hop neighbors find new channel plansto minimize the worst AP’s CM.

A DCA scheme in [8] maximizes a metric based on in-terference graph. The metric models the potential capacityof each AP by taking into account the effects of both co-channel and adjacent channel interference as well as the loadsof neighboring APs. In [3], S. Bae et al. propose a DCAscheme which assigns the channel with the minimum numberof STAs in the overlapping coverage. Each STA periodicallyreports a list of RSS values from its neighboring APs.

All these TPC and DCA schemes, however, do not considerthe physical asymmetry between AP and STA devices, e.g.,

Statistic gathering

Channel plan search

Initial channel assignment

Channel plan sampling TPC

simulation

...

DCA decision

Channel assignment

Transmit power

assignment

Yes

No

Calculate

network cost

Fig. 2. JACPoT framework.

receive sensitivity, antenna gain, etc., which can greatly affectthe performance of WiFi devices. Whereas we propose aTPC and DCA scheme considering the heterogeneity of WiFidevices with different RF parameter set.

III. JACPOT: OVERVIEW

We now present the proposed RRM framework. We assumethat each AP continually collects statistics of RSSs from otherAPs and STAs, noise level, and traffic load in every channelusing a monitoring radio.1An APC periodically gathers thestatistics from the APs under its control and conducts JACPoTbased on the statistics. We call this periodic a run.

We first define a metric, named network cost, for weighingup the network-wide performance for a given channel plan,i.e., the set of channels assigned to APs in the network.Network cost is defined as the sum of APs’ individualmetrics, called residual ICI (RI). More specific descriptionof RI and network cost is presented in Section V. JACPoTis then composed of the following three steps at each run.

• Step 1. Channel plan search: For the optimal channelplan search, we adopt a metaheuristic algorithm annealedGibbs sampler (AGS) [9] which we will discuss specificallyin Section V. For the fast convergence of AGS, we assign theinitial channel plan by a simple greedy search. Specifically,each AP takes turns to choose a channel providing the lowestnetwork cost until no AP can further lower the network costby changing its channel unilaterally. Next, the APC findsthe (near-)optimal channel plan by AGS using the networkcost. The key point is that the network cost is computedusing the expected transmit power levels of APs after TPCoperation. The detailed explanation of the channel plan searchis presented in Section V.

• Step 2. DCA decision: Secondly, we decide whether toassign only transmit power or both transmit power and channeltogether. Due to the temporal disconnection and protocoloverheads such as channel switch announcement (CSA)frame exchange, frequent changes of the operating channelis undesirable. To avoid frequent channel reassignments,

1Many of today’s enterprise WLAN APs are equipped with such a moni-toring radio [4].

therefore, we define two conditions for DCA operation. Thefirst condition is given as

c∗net > α · c̃net (1)

where c∗net represents the optimal network cost found in Step 1,and c̃net represents the network cost of the previous channelplan. α (≥ 1) is a configurable parameter depending on theoverheads of assigning a new channel plan. The APC skipsDCA when the condition above is not satisfied. In addition,to prevent the performance stagnation due to consecutiveskipping of DCA, the second condition is defined as

skip_count = skip_countmax (2)

where skip_count is the number of consecutive runs whichskipped DCAs, and skip_countmax is the maximum allowednumber of skipped DCAs. If either condition is satisfied, theAPC assigns the optimal channel plan to the connected APs.

• Step 3. Channel and transmit power assignment:According to the result of DCA decision, the APC assignschannels and corresponding transmit powers to each AP. IfDCA is skipped, the APC executes TPC with the previouschannel plan and assigns only transmit powers to the APs.The overall operation of JACPoT is described in Fig. 2.

IV. PROPOSED TPC SCHEME

In this section, we describe the proposed TPC scheme indetail. The philosophy of the proposed TPC is to make thecontention domain of an AP, as small as possible, whileguaranteeing stable links between the AP and its client STAs.The contention domain of an AP is the range within whichthe AP’s transmission is sensed by other devices.

We first introduce two conditions for power decrease,namely inner condition and outer condition. Inner condition isthe condition that an AP is able to decrease its transmit power,and outer condition is the condition that an AP is requiredto decrease its transmit power. If an AP satisfies both innercondition and outer condition, it decreases its transmit powerby step size of δP unless it reaches the minimum power level.2

A. Inner Condition

We first define inner condition to guarantee stable linkswithin a cell. Inner condition states that all client STAs shouldbe able to hear the associated AP even with the existence ofhidden interference from co-channel neighbors. It means thatthe transmit power of AP should be kept high enough for allclient STAs to satisfy the SINR requirement for successfulpacket decoding. To this end, AP should know RSSs atits associated STAs from itself. Since there is no standardcompliant protocol for RSS feedback from STAs, the RSS ata STA from an AP should be estimated using RSS of thereverse link, i.e., the RSS at the AP from the STA. We call

2The value of δP depends on AP devices. For example, Cisco Aironet 1140series AP can adjust transmit power by 3 dB, i.e., δP = 3 dB [10].

TABLE IREQUIRED SINR OF 10% FER FOR 802.11n MCSS.

MCS 0 1 2 3 4 5 6 7

SINR 3.56 6.55 8.93 13.03 15.73 20.44 21.69 23.25

this estimated RSS (ERSS). Assuming the channel reciprocity,ERSS is given as

ERSSAP→STA , RSSSTA→AP + PAP − PSTA + ∆comb (3)

where ERSSAP→STA and RSSSTA→AP are the ERSS from APto STA and the average RSS from STA to AP, respectively.PAP and PSTA represent transmit power levels of AP and STA.3

∆comb is the difference of combining gains of AP and STA dueto the different number of antennas. Compared with STAs, APsgenerally have more antennas which yield higher combininggain. All the above values are on the dB scale.

To guarantee a successful packet reception after a unit powerdecrease, the inner condition is expressed as

ERSSAP i→STA t − δP > Stht,k,

∀t ∈ Ti (4)

where Stht,k is RSS threshold in dBm, defined as the required

RSS for STA t to receive packets using target MCS k withthe frame error rate (FER) of 10%, and Ti is a set ofSTAs associated to AP i, respectively. Since inner conditionguarantees a successful packet reception with a target MCS,AP and STAs choose MCSs equal or higher order MCSs thanthe target MCS for data transmissions. By using higher MCSsfor data transmission on average, fewer collisions are expectedthanks to the shorten frame airtime.

A packet using MCS k can be received by STA t with FERof 10%, if the RSS, denoted by St (dBm), satisfies

St − (I +N)dB ≥ SINRk (5)

where SINRk is the required SINR value to get FER of10% using the maximum packet size4 and MCS k, and itis computed by the method presented in [12] as Table I.I and N represent interference and noise power, respectively.Therefore, RSS threshold of STA t using MCS k is given by

Stht,k , SINRk + (I +N)worst,dB + λt,p (6)

where (I +N)worst represents the power of interference plusnoise in the worst case. λt,p is a margin for signal fluctuationdue to multipath fading, and it is added to cover up to the(1− p)th percentile of RSS values.

We expect that the worst interference to a STA happenswhen its associated AP cannot sense the interference by anarrow margin, because the AP would not transmit with theexistence of an interference louder than the CCA threshold.If we only consider interference from co-channel APs,5 STAs

3The RF parameters of STA is assumed to be known by the APC, e.g.,through one-time device registration with the enterprise WLAN system.

4The maximum MSDU size for 802.11n WLAN is 7,935 bytes [11].5APs transmission occupy the air during most of time, because downlink

traffic dominates the total traffic in general 802.11 WLANs [13].

generally hear signals from neighbor APs with smaller RSSthan its associated AP.6 Therefore, the worst case interferenceis given by the CCA threshold of the AP. Consequently, theRSS threshold of STA t is rewritten as

Stht,k = SINRk + (Cth + NFt)dB + λt,p (7)

where Cth is the AP’s CCA threshold and NFt is STA t’snoise floor, both on the linear scale.

On the other hand, To determine the value of λt,p, each APcollects RSS statistics and find the top p RSS value for eachSTA. For example, if the average RSS and top 95% from aSTA are −60 dBm and −70 dBm, respectively. λt,95% is givenby −60 − (−70) = 10 dB. p can be adaptively determinedaccording to traffic type and load.

B. Outer Condition

Secondly, we define outer condition to reduce contentiondomains of APs as much as possible. Outer condition statesthat an AP needs to reduce its transmit power, if any of itsco-channel neighbor APs hears it as louder than the CCAthreshold. Thus, outer condition is given as

RSSi,j > Cthj ,

∃j ∈ Ai (8)

where RSSi,j and Cthj represent the average RSS from AP i

at AP j and AP j’s CCA threshold, respectively, and Ai is aset of AP i’s co-channel APs. We only consider ICI betweenAPs for the same reasons as in footnote 6.

V. PROPOSED DCA SCHEME

In this section, we propose a DCA scheme which is jointlyexecutable with the proposed TPC scheme. Note that theproposed TPC scheme guarantees that links between an APand its client STAs are robust to hidden interference. Thepurpose of the proposed DCA scheme is to minimize theperformance degradation due to contention among co-channelcells. The bottom line is assuming that the proposed TPChas been done for each channel plan, and finding the optimalchannel plan which has the lowest degree of contention.

A. Residual Inter-cell Interference

We borrow the concept of contention factor (CF), a DCAmetric proposed in [15], to estimate the degree of contention.CF assumes that performance degradation due to contention isproportional to the gap between the current power level and thedesirable power level (DPL). DPL is the maximum allowedtransmit power for an AP in order not to cause exposedinterference to co-channel APs.

In this work, we modify CF into a new metric which isnamed RI. The modification is twofold. First, RI expects thedegree of contention after the proposed TPC is applied, while

6In the worst case, the distance between an AP and a neighbor AP is 2×longer than the distance between the AP and its client STA. With pathlossexponent of 3.5, the pathloss difference due to the distance difference is10.5 dB, which can be compensated by the different device characteristics,e.g., AP’s higher antenna gain or handgrip loss of mobile STAs [14].

CF measures degree of contention with the current power.Second, RI has a form of log function, while CF is a linearfunction. To be specific, we newly define two terms; desirablepower decrease (DPD) and allowed power decrease (APD).DPDi,j represents how much power AP i should decrease notto satisfy outer condition by AP j.

DPDi,j , min([RSSi,j − Cth

j

]+(9)

where [x]+ = max(0, x). Next, APDi represents how muchpower AP i can decrease while satisfying inner condition.

APDi , min

P̃i − Pm,i,

mint∈Ti

(ERSSi,t − Sth

t,k

)δP

× δP

(10)

where P̃i and Pm,i are AP i’s current and minimum transmitpowers in dBm, respectively, and bxc is the greatest integerless than or equal to x. Then, the performance degradationdue to AP i is analyzed by DPD and APD as follows.

1) If APDi > DPDi,j , ∀j 6= i, AP i can lower its transmitpower until no neighbor APs can sense its signal, i.e., noneighboring APs contend with AP i. Thus, AP i causesnegligible performance degradation in the network.

2) If APDi ≤ DPDi,j , ∃j 6= i, AP i is sensed by AP j afterlowering transmit power by the proposed TPC, and AP jcontends with AP i. In this case, the degree of contentionat cell j depends on ∆i,j , DPDi,j − APDi for tworeasons. Firstly, more STAs in cell j can sense a signalfrom AP i and contend with the AP as ∆i,j increases.Secondly, the probability that AP and STAs in cell jsense a signal from AP i increases as ∆i,j increases.7 Theeffects of increasing ∆i,j on the performance degradationfrom contention becomes less significant as ∆i,j becomeslarger, and eventually converges to a certain level.

We capture this relationship between the performance degra-dation and ∆i,j by modeling it with a log function, which isnon-decreasing and concave. Thus, RIi is defined as

RIi , τi∑j∈Ai

log(

1 + [DPDi,j − APDi]+)

(11)

where τi is AP i’s traffic load.8 Network cost is then definedby the sum of N AP’s RI as

cnet ,N∑i=1

RIi. (12)

7Due to the multi-path fading, a signal from AP i may not be sensed byAP j temporarily, although RSSi,j ≥ Cth

j . If ∆i,j is large enough, thedevices in cell j always sense a signal from AP i.

8Traffic load can be defined in various ways, e.g., number of associatedSTAs, traffic source rate, or channel occupancy time. In this work, we definethe traffic load as the transmission time of AP i divided by total time.

Algorithm 1 JACPoT1: ch← GreedySearch . Initialize channel plan by greedy search in Section III

(ch: Channel plan)2: (c∗net, p)← FINDNETCOST(ch) . Initialize optimal network cost (c∗net)

(p: Transmit power plan)3: t← 1

4: while t ≤ tmax do . For each iteration t < tmax

5: T ← 1log2(1+t)

. Update temperature according to cooling schedule

6: for i← 1...n do . For all APs (n: # of APs)7: for all chi ∈ 1, 2, · · ·,m do . For all channels (m: # of channels)8: ch← (ch1, · · ·, chi, · · ·, chN )

9: (cnet(chi), p)← FINDNETCOST(ch)

10: µi(chi)←(e−

cnet(chi)T

)/(∑ch′

i∈{1,2,···,m} e

−cnet(ch

′i))

T

)11: end for12: pick chi ∼ µi(chi) . Randomly choose chi according to µi(chi)

13: update (ch, p) using chi

14: end for15: if cnet < c∗net then . Update optimal plans according to cnet16: (ch∗, p∗, c∗net)← (ch, p, cnet) (ch∗: Optimal channel plan,17: end if p∗: Optimal transmit power plan)18: t← t+ 1

19: end while

20: function FINDNETCOST(ch)21: for i← 1...n do . For all APs in the network22: APDi ← compute APD using (10)23: for j ← 1...n do24: DPDi,j ← compute DPD using (9)25: RIi ← compute RI using (11)26: end for27: pi ← min(maxj DPDi,j , APDi)

28: end for29: cnet ←

∑RIi

30: return (cnet, p) . p = (p1, p2, ..., pn)

31: end function

B. Annealed Gibbs Sampler (AGS)

Now, we find a channel plan minimizing the above-definednetwork cost. However, the number of possible channel plansgrows exponentially with the number of APs in the network.For example, a network composed of 10 APs with 4 channelshas 410 ≈ 106 possible channel plans. In other word, if wesearch exhaustively, the search space size is about 106. Toreduce the search space and make this optimization morefeasible, we use AGS to find the optimal channel plan.

AGS combines the notion of Gibbs sampler and simulatedannealing [16]. To put it briefly, each node updates its state,i.e., s, according to jumping probability, which is given as

µ(s) =e−ε(s)/T∑

s′∈S e−ε(s′ )/T

, (13)

where S is a set of possible states. Energy denoted by ε(s)is a measure of each state’s optimality, where a state withsmaller energy is more desirable. In JACPoT, a state andenergy correspond to a channel plan and the network costof the channel plan, respectively.

Temperature, denoted by T , controls the sensitivity ofjumping probability to the energy. If temperature is high, thejumping probability of a state less depends on the state’senergy. For convergence to the optimal state, the temperature

0 2000 4000 6000 8000 10000

25.2

25.4

25.6

25.8

26

Iteration ( t )

Op

tim

al n

etw

ork

co

st

Exhaustive search (1050 s)Random sampling (0.38 s)AGS T = 10

1+log2t(20.43 s)

AGS T = 11+log2t

(20.25 s)

AGS T = 11+log10t

(20.10 s)

Fig. 3. Convergence of search algorithms.

is updated according to cooling schedule for each iteration t.We adopt a cooling schedule which is inversely proportionalto a logarithmic function of t [17] as T = T0

log(1+t) .Fig. 3 shows the convergence of optimal network cost

from exhaustive search, random sampling, and AGS withdifferent parameters. The numbers in the legend representaverage calculation time for 10,000 iterations through Intel(R)Core(TM) i5-4670 CPU @ 3.40 GHz. For 20 networks eachcomposed of 13 APs and 100 randomly distributed STAs, theoptimal network cost according to t is averaged out.

Compared with random sampling, the result of AGS algo-rithms, especially the AGS with logarithm base 2 and T0 = 1,rapidly converges to the optimal value from exhaustive search.In addition, average calculation time of AGS algorithms isabout 20 seconds, and smaller than that of exhaustive searchby several orders of magnitude. Since the calculation time ofAGS grows proportionally (not exponentially) to the numberof APs, and is readily controllable by limiting the maximumnumber of iterations, it is possible to find the (near-)optimalnetwork cost within a reasonable time. The detailed operationof JACPoT is described in Algorithm 1.

VI. PERFORMANCE EVALUATION

In this section, the performance of JACPoT is evaluatedvia ns-3 simulation [18]. For a realistic simulation, we firstmodified TGn pathloss model (developed to evaluate 802.11nproposals) based on the measurement results in real officeenvironments with many cubicles and glass walls. As a result,we use a higher pathloss exponent and rapidly decreasingsignal strength at long distance due to the penetration loss. Tocapture the physical asymmetry between AP and STA devices,in addition, we adopt a WiFi simulation model proposedin [14]. We consider an enterprise wireless network with twotier hexagonal cell deployment with 19 APs as shown inFig. 4. Four non-overlapping available channels are assumed,and Minstrel rate adaptation algorithm [19] using 802.11ndata rates with a single spatial stream is adopted. Detailedsimulation environment is described in Table II.

A. DCA Metric

First, we compare the proposed metric, i.e., RI, with existingmetrics to see how correctly it measures the network perfor-

Fig. 4. Two tier hexagonal cell topology with regular channel plan.

TABLE IISIMULATION MODELS AND PARAMETERS.

Model & parameter Description

Transmit power −1 ∼ 20 dBm (AP), 12.5 dBm (STA)CCA threshold −82 dBmAntenna gain 3 dB (AP), 0 dB (STA)Handgrip loss [14] −6.2 dB (STA)Noise floor 3.5 dB (AP), 9.4 dB (STA)Pathloss model 28.8logd+ 46.7, d ≤28.5 m

46.1logd+ 20.5, otherwiseFading model Jakes model (doppler velocity= 0.1 m/s)

mance. For a given topology with a random channel plan, wemeasure both sum throughput and sum of each metric, i.e.,RSS, COT, Load, and RI. Specifically, RSS is given similarto the metric proposed in [5] as mrss

i =∑j∈Ai

RSSj,i. COTis the metric proposed in [7], and is given by mcot

i = COTi ·∑j∈Ai

RSSj,i, where COTi is the channel occupancy timemonitored by AP i. Load is the metric proposed in [6], and isgiven by mload

i = ρi +∑j∈Ci(1)

ρj +∑

(m,n)∈Ci(2)ρmρn,

where ρi and Ci(k) are the effective channel utilizationand class-k interferers for AP i, respectively. The inter-APdistance (dAP) is 10, 15, 20 m, and the number of STAs in acell is a uniform random variable over [1, Nmax].

- RI shows strong correlation with network throughputFig. 5 shows the correlation between the throughput and the

metrics for 100 random channel plans. The throughput andthe sum metric values are normalized to the aggregate sourcerate and the maximum value of sum metric, respectively. Thediagrams show that RI achieves higher (negative) correlationwith the throughput compared with other metrics. In addition,RI is more discriminative in the sense that the values of summetric are more widely scattered in x-axis.

- The accuracy of RI increases with more number of STAsTable III shows the average correlation coefficient of each

metric for different Nmax. RI yields very strong correlationespecially when there are many STAs since the performancedegradation modeling in Section V-A fits better with a largenumber of STAs. Therefore, our scheme is more suitable fordense environments.

B. Performance of JACPoTNext, we randomly distribute 100 STAs in the network

with 19 cells. All STAs have TCP traffic of 0.8 Mbps fordownlink (DL) and 0.2 Mbps for uplink (UL). 10% of totalSTAs have VoIP traffic using G.711 codec with inter-packetinterval of 20 ms. The sum throughput of DL and UL traffic

0.99 0.995 10.4

0.6

0.8

1

Normalized sum metric

No

rma

lize

d t

hro

ug

hp

ut

(a) RSS

0.99 0.995 10.4

0.6

0.8

1

Normalized sum metric

No

rma

lize

d t

hro

ug

hp

ut

(b) COT

0.4 0.6 0.8 10.4

0.6

0.8

1

Normalized sum metric

No

rma

lize

d t

hro

ug

hp

ut

(c) Load

0.7 0.8 0.9 10.4

0.6

0.8

1

Normalized sum metric

No

rma

lize

d t

hro

ug

hp

ut

(d) RI

Fig. 5. Correlation between normalized aggregate TCP throughput and summetrics: dAP= 10 m, Nmax= 10.

TABLE IIICORRELATION COEFFICIENTS BETWEEN EACH METRIC AND THROUGHPUT.

Nmax RSS COT Load RI

10 −0.39 −0.46 −0.77 −0.9220 −0.36 −0.39 −0.51 −0.9530 −0.38 −0.39 −0.35 −0.97

as well as and R-score, which is given by the minimum ofR-scores for DL and UL, are measured for 20 s. For eachdAP= 10, 15, 20 m, we repeat the simulation 10 times withdifferent random seeds. As comparison schemes, we chooserandom channel plan with fixed transmit power (Rnd./Fixed),and regular channel plan as shown in Fig. 4 with Cisco TPCversion 1 (Reg./Cisco).

- 97% STAs achieve full throughput in both UL and DLFigs. 6(a) and 6(b) show the CDF of STAs’ normalized

throughput, i.e., throughput normalized to the source rate, andR-score. The results for all dAPs are plotted in the samegraph, and mean (m) and Jain’s fairness index (J) of eachscheme are shown in the legend. Compared with Rnd./Fixedand Reg./Cisco,9 JACPoT significantly enhances throughput.The ratio of STAs not fully receiving and transmitting thegiven traffic is 0.03 in JACPoT, while they are 0.54 and 0.47in Rnd./Fixed and Reg./Cisco, respectively. On the other hand,JACPoT drops the ratio of STAs with R-score lower than70 by 52% compared with Reg./Cisco. On average, JACPoTachieves 14% performance gain over Reg./Cisco for both dataand VoIP traffic, and also significantly enhances the fairness.

- DL throughput gain using higher order MCSsError statistics shown in Figs. 6(c) and 6(d) explain the

performance gain. For DL transmission, the ratios of framereception error due to low RSS, i.e., channel errors, and the

9Rnd./Fixed is the average of 2, 8, 14, 20 dBm fixed powers with randomchannel plans. Similarly, Reg./Cisco is the average result of Cisco TPC using−50, −60, −70, −80 dBm TPC thresholds with the regular channel plan.

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

Normalized throughput (Mbps)

CD

F

Rnd./Fixed (m = 0.82, J = 0.92)

Reg./Cisco (m = 0.87, J = 0.95)

JACPoT (m = 0.99, J = 1.00)

(a) CDF of normalized throughput

0 10 20 30 40 50 60 70 80 900

0.2

0.4

0.6

0.8

1

R−score

CDF

Rnd./Fixed (m = 77, J = 0.94)

Reg./Cisco (m = 73, J = 0.91)

JACPoT (m = 83, J = 0.98)

(b) CDF of VoIP R-score

0

0.05

0.1

0.15

0.2

10m 15m 20m 10m 15m 20m 10m 15m 20m

Pa

cke

t e

rro

r ra

te Channel error

Collision error

JACPoTReg./CiscoRnd./Fixed

(c) DL error rate

0

0.05

0.1

0.15

0.2

10m 15m 20m 10m 15m 20m 10m 15m 20m

Pa

cke

t e

rro

r ra

te Channel error

Collision error

JACPoTReg./CiscoRnd./Fixed

(d) UL error rate

Fig. 6. Performances of TPC/DCA schemes: dAP= 10, 15, 20 m, k = 3, p = 95%.

error due to interference, i.e., collision errors, are comparable.Even with a slightly higher channel error rate compared withReg./Cisco, JACPoT yields greater throughput in DL becausehigher order MCSs are used on average. In addition, JACPoTkeeps DL channel error rate under 10% regardless of dAP asthe proposed TPC guarantees.

- UL collisions are reduced by 36%On the other hand, collision errors are dominant in UL,

since STAs have more hidden terminals due to handgrip lossand small antenna gain. By reducing ICI, JACPoT reduces ULcollision errors by 36% compared with Reg./Cisco. Therefore,VoIP performance which is generally limited by UL frameerrors is enhanced by JACPoT.

VII. CONCLUDING REMARKS

In this paper, we have proposed JACPoT, a centralizedand jointly executable TPC and DCA scheme for enterpriseWLAN. While most existing schemes do not consider thephysical asymmetry between AP and STAs devices, JACPoTis designed by focusing on the network environment with theasymmetry. We confirm the feasibility of JACPoT based onAGS, and examine the accuracy of the proposed DCA metric.Finally, the simulation results demonstrate JACPoT enhancesboth data and voice traffic delivery by reducing collisions andpromoting higher data rate usage. As future work, we areplanning to extend our scheme considering channel bonding.

ACKNOWLEDGMENT

This work was supported by ICT R&D program ofMSIP/IITP. [B0126-15-1017, Spectrum Sensing and FutureRadio Communication Platforms]

REFERENCES

[1] Status of IEEE 802.11ax. [Online]. Available: http://www.ieee802.org/11/Reports/tgax update.htm.

[2] W. Li et al., “Achieving proportional fairness via AP power control inmulti-rate WLANs,” IEEE Trans. Wireless Commun., vol. 10, no. 11,pp. 3784–3792, Nov. 2011.

[3] S. Bae et al., “Self-configuration scheme to alleviate interference amongaps in IEEE 802.11 WLAN,” in Proc. IEEE PIMRC, Sydney, Australia,Sep. 2012.

[4] Cisco Systems, Cisco Wireless LAN Controller Configuration Guide:Software Release 7.0. [Online]. Available: http://www.cisco.com/en/US/docs/wireless/controller/7.0/configuration/guide/c70.html.

[5] R. Akl and A. Arepally, “Dynamic channel assignment in IEEE 802.11networks,” in Proc. IEEE PORTABLE, Orlando, USA, Mar. 2007.

[6] K. K. Leung and B.-J. Kim, “Frequency assignment for IEEE 802.11wireless networks,” in Proc. IEEE VTC, Orlando, USA, Oct. 2003.

[7] G. Kazdaridis et al., “Novel metrics and experimentation insightsfor dynamic frequency selection in wireless LANs,” in Proc. ACMWiNTECH, Las Vegas, Nevada, USA, Sep. 2011.

[8] E. G. Villegas et al., “Frequency assignments in IEEE 802.11 WLANswith efficient spectrum sharing,” Wireless Commun. Mobile Comput.,vol. 9, no. 8, pp. 1125–1140, 2009.

[9] B. Kauffmann et al., “Measurement-based self organization of inter-fering 802.11 wireless access networks,” in Proc. IEEE INFOCOM,Anchorage, Alaska, USA., May 2007.

[10] Cisco Aironet 1140 series access point. [Online]. Available:http://www.cisco.com/en/US/prod/collateral/wireless/ps5678/ps10092/datasheet c78-502793.html.

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

[12] D. Qiao, S. Choi, and K. G. Shin, “Goodput analysis and link adaptationfor IEEE 802.11a wireless LANs,” IEEE Trans. Mobile Comput., vol. 1,no. 4, pp. 278–292, Oct. 2002.

[13] A. Gupta, J. Min, and I. Rhee, “Wifox: Scaling wifi performance forlarge audience environments,” in Proc. ACM CoNEXT, Nice, France,Dec. 2013.

[14] S. Yoo et al., “Toward realistic WiFi simulation with smartphone“Physics”,” in Proc. IEEE WoWMoM, Sydney, Australia, Jun. 2014.

[15] S. Kim et al., “A novel metric for dynamic channel assignment inenterprise wireless networks,” in Proc. ICUFN, Da Nang, Vietnam, Jul.2013.

[16] P. Bremaud, Markov Chains, Gibbs Field, Monte Carlo Simulation andQueues. Springer-Verlag, 1999.

[17] H. Szu and R. Hartley, “Fast simulated annealing,” Physics Letters A,vol. 122, no. 34, pp. 157–162, 1987.

[18] Network Simulator ns-3. [Online]. Available: http://www.nsnam.org/.[19] MadWiFi, Minstrel rate control. [Online]. Available: http:

//madwifi-project.org/browser/madwifi/trunk/ath rate/minstrel.