A Study on Dynamic Load Balance for IEEE 802.11b Wireless LAN

by Ioannis Papanikos and Michael Logothetis

EE 228a - Fall 2003 Dennis Chang

Introduction• IEEE 802.11 defines two types of wireless networks for

different communication needs:• Independent Basic Service Set (IBSS) or Ad Hoc mode:

• Limited in its range – all stations need to “see” or “hear” each other

• Peer-to-peer communication among the stations

• No fixed wired infrastructure for stations to communicate with each other

Introduction• Extended Service Set (ESS) or Infrastructure mode:

AP AP

Wire station

Wireless station

(WS)

• Wireless Stations are associated to an Access Point and through AP communicate with other wire or wireless stations

Introduction• WS can be supported by more than 1 AP in the same

region• Operation functions of the AP (Channel, ESSID, WEP,

etc.) can differ among the AP while they support the same network

• For roaming function, all AP in the same region need the same key functions as ESSID and WEP, but can operate in different channels (14, 11 in U.S., 3 non-overlapping)

• Maximum number of associated stations per AP is 2007• No function specifying the AP selected by a WS

Load Balance Problem• One AP may support many WS, while some neighbor AP

may support few or no stations

AP 1AP 2

AP 3

• this asymmetry in load between AP causes a high probability of packet loss in AP 1 and thus an overall network degradation compared to AP 2, 3

• Can be avoided by balancing the number of associated stations among AP

Load Balance Problem

AP 1AP 2

AP 3

AP 1AP 2

AP 3

• Whether applying a WS distribution algorithm results in a load balance for each AP depends on the nature of the wireless LAN

• Problem becomes more difficult due to dynamic network topology changes when the WS are roaming around

• Would be impractical to consider a station rearrangement after every new connection/disconnection

Load Balance Problem• Classical Approach• Proposed Dynamic Load Balance Algorithm – 3 levels:

– AP Channel Autoselection Level – tries to best distribute the APs to the available channels

– Station Join Decision Level – method by which the WS select the AP to associate with

– Link Observation Level – determines when the WS leaves its associated AP and the roaming function is performed (seeks a newAP)

• Results

Classical Approach• Currently, most manufacturers implement this procedure

for association of a WS to an AP:– WS scans the available channels of each AP in the region (active

scan with Probe Request)– Listens to the Beacon or Probe Response Frames– WS stores the Received Signal Strength Indicator (RSSI) of the

received frames and other info. (ESSID, encryption (on/off), etc.)– After scanning, WS selects AP with the maximum RSSI, given

that the AP covers the other requirements (ESSID, WEP) also– WS leaves AP when the RSSI falls under a predefined threshold

Classical Approach• Procedure is based on the hope that the quality of service

of the selected AP is the best• But this typically results in many stations being connected

to a few AP, and other neighboring AP being idle, as mentioned before

• Overloaded AP leads to performance degradation• A proper load balance algorithm needs to consider the

status of each AP and its already associated WS, in order to associate new WS to an AP

Dynamic Load Balance – Level 1• AP Channel Autoselection Level

– At start-up phase of each AP, the AP is informed of the existence of other APs in the same region by the Inter Access Point Protocol (IAPP), which transfers information proving that the AP service the same LAN.

– AP active scans channels to discover which other AP are neighbors– Also it discovers the operation channels of the neighboring AP– Starts using the operation channel where the interference from

neighboring AP is minimized– This level is initialization step to normalize operation conditions of

the network– At the time of paper (2001) there is no AP with such initialization

level, due to difficulties in implementation– So initial operation channel is chosen to be 1, 6, or 11 in U.S.

Dynamic Load Balance – Level 2• Station Join Decision Level

– WS sends a Probe Request to all channels in order to localize the AP, AP answer with Probe Response containing extra info. than already defined in protocol:

• Ni : Number of stations associated to AP i• Si : RSSI value of the Probe Request received by AP i• Mi : Mean RSSI value for the set of stations associated to AP i

(including RSSI for the new WS)– After the scanning procedure, WS determines the best AP to

associate with, and sends an Association Request to that AP

Dynamic Load Balance – Level 2

Wireless Station Access Point iProbe Request

Probe Response (Ni, Si, Mi)

Save Ni, Si, Mi

Association Request

Determine best AP to join

AP i

Dynamic Load Balance – Level 2• So how does the WS determine the best AP to associate

with? Authors provide this formulation:– Select AP that maximizes the weighted function Wi = Di * Pwi * Pi

– Di denotes the difference between RSSI received from AP I, Si, and the mean RSSI

– Pwi is a weight proportional function that considers not only differences from mean value but also the absolute value of the mean value

– Pi is the weight proportional to the number of already associated WS to AP i

≤−≥+

=0/10/1

iii

iiiwi

DifSMDifSM

P

iii SMD −=

region in the AP ofnumber total theisn where,

0∑

=

= n

jj

ii

N

NP

Dynamic Load Balance – Level 2• Corrected formulation (perhaps):

– Select AP that Maximizes Wi = -Di * Pwi * Pi– Or Minimizes Wi = Di * Pwi * Pi– Where

∑−

=

=

≤−≥+

=

−=

1

0

0/10/1

n

jj

ii

iii

iiiwi

iii

N

NP

DifSMDifSM

P

MSD

Dynamic Load Balance – Level 3• Link Observation Level

– Determines when the WS leaves the AP and the roaming function is performed (seeks a new AP)

– Each AP updates the mean RSSI value, Mi, and the number of associated stations, Ni, in each Beacon or Probe Response Frame

– WS periodically probes the AP and updates RSSI in the side of APi, Si, Mi, and Ni, or monitors the Mi and Ni through the Beacon frames

– Each time the WS detects a worse case than that calculated previously, it increases a Handover Counter (HC)

– Once HC reaches a predefined threshold, the WS seeks a new AP– Threshold depends on manufacturing characteristics of WS and AP

(sensitivity, transmission power)

Protocol Modifications• New fields are needed in Beacon and Probe Response

frames for transmitting extra information for the algorithm– Ni – number of associated stations (Beacon, Probe Response)– Si – RSSI of the incoming Probe Request (Probe Response)– Mi – mean RSSI for the associated stations (Beacon, Probe

Response)

• AP must be able to activate active or passive scanning procedures

• Both the WS and AP sides need to support the algorithm

Results• Algorithm tested using Charriot Test Utility, with 3 APs

and 30 WS which transferred data files of 12 MB to and from an Ethernet network– Traffic load conditions were the same for all WS– Results showed balanced distribution (symmetry) of number of

WS to the AP and improvement in overall network performance, especially in the absence of hidden stations

– But with many hidden stations, “results were not satisfactory”– Also unsatisfactory results with unbalanced traffic to or from the

WS– Hidden stations caused a lot of problem because of the large

number of collisions when one or more WS cannot hear other WS that transmit

Results– Algorithm needs to account for the hidden stations– Need good estimation of the number of hidden stations– Also, to deal with large traffic variations among the WS and the

resource availability of the AP, we want to include real time measurements of the resource availability and frame error rate in the Station Join Decision Level (level 2)

Another DLBA• “Dynamic Load Balance Algorithm (DLBA) for IEEE

802.11 Wireless LAN” by Shiann-Tsong Sheu and Chih-Chiang Wu (1999)

• Problem description, problem formulation, and algorithm are very similar

• Will briefly discuss this approach..

Load Balance Problem• M STAs (mobile stations) and N APs (access points)• Each AP/STA can access one channel at a time• STA can select 1 AP to associate with and assume each AP

is capable of supporting at least M STAs• Sx : set of STAs connected to APx

• SNx : number of stations in Sx = |Sx|• Rx(y) : RSSI when APx receives packets issued from STAy

• ARx : average RSSI value of set Sx = average RSSI between any STA in Sx and APx

x

Sxyx

xSN

yRAR

∑ ∈=)(

Load Balance Problem• VAR : variance of ARs• VSN : variance of SNs• Dynamic Load Balancing Problem:

– Given a WLAN with a number of STAs and APs and each STA selects 1 AP, find a scheme such that

• (1) The average RSSI is maximized, and• (2) Both VAR and VSN are minimized

Load Balance Problem• Can be represented as a bipartite graph except that it is a

dynamic assignment problemAPs

Y1 Y2 Y3X1 7 5 1

STAs X2 3 8 4X3 6 3 5X4 2 9 9

X1

X2

X3

X4

Y3

Y2

Y1

9

95

364 8

31 5

7

2

STAs

APs

• Static algorithms like Hungarian Algorithm have been proposed to solve the basic assignment problem, but not suited for DLBP since the process of STAs joining and leaving is dynamic, and it’s impractical to rearrange all stations’ assignment whenever a new station joins or leaves

Signal Strengths

Load Balance Problem• Rmax : normalized maximum RSSI value which can be

received and estimated by WLAN adapter• Like in previous algorithm, AP includes additional info. in

Probe Response frame:– ARx’ : new average RSSI which includes STAy with set Sx

– Rx(y) : RSSI when APx receives probe request from STAy

• Major reference value in proposed algorithm:– Difference tells affect of station y joining set x

1)()(

'+

+= ∑ ∈

x

Sxzxx

xSN

yRzRAR

')()( xxx ARyRyD −=

Load Balance Problem• STAy prefers selecting the AP which has the maximum D(y) among

all APs• Based on this, the average RSSI in sets may perform very close to the

classical approach (i.e. we are maximizing average RSSI), but it still doesn’t guarantee that stations are equally distributed to different sets

• Solution:– Stations whose RSSI are less than the new average RSSI, AR’ are forced

to change into another set– Handoff process decreases performance, so maintain a holding counter

HC. Each time a new station joins the set, observe new average RSSI and and compare with its own RSSI value. Increment HC if necessary

– Once threshold reached, leave set and become a new station, reset HC– Similar to Link Observation Level (level 3) of previous algorithm

Load Balance Problem• Since handoff process only when station’s HC reaches a

threshold, this progress will eventually terminate• In effect, all stations will be rearranged into a relatively

better condition than before, so there’s improvement• With a smaller threshold, a better load balance will be

obtained but with more thrashing

Load Balance Problem• Another problem:

– If a new station only considers the difference between R(y) and AR’, it may choose a worse AP

– Example: STAy should join APa, not APb

50

80

90

1015

30

0

10

20

30

40

50

60

70

80

90

100

ARa’ Ra(y)ARa ARb ARb’ Rb(y)

Db(y) = 15Da(y) = 10

Load Balance Problem• Solution: Consider the AR’ in decision problem. Introduce a

proportional weighted function P(y):

• Weight of a station STAy connecting to APx:

• When a station wants to join, it calculates weights of APs and joins one with maximum weight

<−

≥+=

0)(,'1

0)(,'1)(

max

max

yDifRAR

yDifRAR

yPx

x

xx

x

)(*)()( yPyDyW xxx =

Flowchart of AP

APx

Receive a Probe

Request from STAy

Calculate ARx(y)’ and Rx(y)

Send Probe Response for

STAy with ARx(y)’ and Rx(y)

Yes

No

Flowchart of STANew Station STAy

Set Channel = 1

Channel > Maxchannel

Send Probe Request

on Channel

Calculate weights and select

the best AP to join

Receive a Probe

Response from APx

Old Stationyes

yes

no no

no

yes

Timeout

Channel++

Record ARx(y)’

and Rx(y)

Simulation Results• Algorithm said to be Load balancing if all APs have the same number

of members and the average RSSI is maximized• Average RSSI value in DLBA is not as high as in classical approach,

but the load of APs with DLBA is more balanced• VSN (variance of set size) of DLBA increases slightly as network size

increases, and VAR (variance of average RSSI) decreases slightly �

proposed algorithm has ability to fairly distribute stations into all APsand guarantee near optimal average RSSI no matter how many stations

• Authors also point out that RSSI is not necessarily a reliable indication of performance, and that a good estimation method should includequality of transmission which is often measured by frame error rate (FER)

NM

Questions/Comments