2014 IEEE 28th International Conference on Advanced Information Networking and Applications

Application-Aware Fast Dormancy in LTE

Jacques Bou Abdo Imad Sarji Imad H. Elhajj Ali Chehab Ayman Kayssi
American University of Beirut
Department of Electrical and Computer Engineering
Beirut 1107 2020, Lebanon
E-mails: { jb48, ies02, ie05, ayman, chehab}@aub.edu.lb

Abstract- Two Radio Resource Control states have been again from LTE_ACTIVE long DRX to LTE_IDLE. If the
proposed in LTE and implemented to ensure low UE power two timer values are not tuned properly, the network can be
consumption and high network resource availability. Transiting congested with extra unused signaling, which causes severe
between these two states optimizes network performance if tuned battery degradation. It will be shown in section 4 that no
properly. Currently, a UE switches from the LTE_ACTIVE state
to the LTE_IDLE state after a pre-configured static inactivity
constant LTE_IDLE timeout is suitable for all users, thus
duration. This paper seeks to demonstrate that no static timeout is fixed timeout values at the network level will neither
optimal for all users at all times. In addition, a user-level dynamic optimize network performance nor battery consumption. We
decision algorithm is proposed to have fine-grain user level propose in this paper a dynamic LTE_IDLE timeout for each
optimization. Since achieving better efficiency is related to context user based on traffic generation rate (establishing/receiving
awareness, we present a solution that allows the UE to auto-learn legacy and application calls in addition to keep-alive
its traffic behavior. The dynamic algorithm was applied to five messages resulting from installed applications), which will
different user load scenarios of combined application and legacy help the network with reduced signaling and will allow the
traffic, and the results showed that we are able to attain power UE to achieve lower battery utilization.
savings of up to 30% when compared to the fixed timeout case.
This paper is divided as follows: Section 2 lists previous
I. INTRODUCTION mechanisms to minimize signaling traffic. Section 3 presents
background information about EPC states and fast
In early third generation cellular (3G) networks, and dormancy. Section 4 describes the mathematical analysis of
before the introduction of cell_PCH (Cell Paging CHannel), the LTE_IDLE threshold specifying when it is beneficial for
the user equipment (UE) had to be either in dedicated or in the UE to transit to LTE_IDLE. Section 5 shows the
idle state. The dedicated state results in fast response during algorithm developed to decide if each UE should transit to
call setup; however, it causes high power consumption and LTE_IDLE or LTE_ACTIVE long DRX. Section 6 shows
hence leads to short battery life [10]. On the other hand, the experimental values based on popular always-online
idle state has much lower power consumption, but results in applications to identify the frequency of sent keep-alive
extra signaling and delay during call setup [8-9]. While in messages. Section 7 proposes an auto-learning algorithm that
dedicated state, only two signaling messages are needed to allows the UE to infer the keep-alive pattern of the installed
establish a legacy call, while 38 signaling messages are applications. Section 8 shows a comparison between our
needed in the idle state. Smartphones have high power proposed decision algorithm and different fixed timeouts,
consumption due to high-resolution screens and complex and section 9 concludes the paper.
processing needed for their sophisticated operating systems
II. RELATED WORK
(OS) and installed applications. Therefore, smartphone
vendors started implementing a fast transition to idle state as
Smartphone proliferation puts pressure on cellular service
a procedure to increase battery life. This, as a result,
providers to operate under constrained resources including
increased the amount of signaling traffic.
radio capacity, network processing overhead, in addition to
Cell_PCH was introduced in Release 99 of the 3GPP
handset limited battery. The solutions proposed to take most
standards as an intermediate tradeoff state having acceptable
advantage of the available resources can be categorized into
power consumption, minimized delay, and minimized
three approaches:
number of signaling messages during call setup. Similar to
cell_PCH, fast dormancy (discussed in section 2) is also 1) Network Inactive Timers
considered as a trade-off solution. Radio management has become a critical issue for
EPS (Evolved Packet System) proposed only two states: operators to satisfy the increasing market demand. Huge
LTE_ACTIVE and LTE_IDLE, where LTE_ACTIVE has efforts have been put to optimize inactive timers in order to
long DRX (Discontinuous Receive) and short DRX options release network resources when not needed. In [9], the
[1-7]. LTE_ACTIVE long DRX replaces cell_PCH, where authors studied the performance of web and voice
the UE transits from LTE_ACTIVE to LTE_ACTIVE long applications for a fixed timeout assuming low page
DRX after a “drx-InactivityTimer” period of silence. After reactivation probability. The results showed that setting the
an “LTE_IDLE timeout” period of silence, the UE transits tail time to [10-15] seconds gives an acceptable outcome

1550-445X/14 $31.00 © 2014 IEEE 194

DOI 10.1109/AINA.2014.28
from service quality and signaling overrhead perspectives. retransmission procedure willl take place after HARQ RTT,
However, shorter timers will cause degraddation of the user’s which is a parameter configuured by the network for each UE
experience and higher overhead on the neetwork side. Using and represents the number of sub-frames after which the
fixed timeouts is beneficial for specific caases, but setting the failed retransmission by a speecific UE is expected.
threshold is very challenging and might leead to a drastic loss Providing an efficient sccheduling power consumption
when used in other scenarios. In [10], the authors studied the model for the connected devices
d has always been high
impact of tail timers for different 3G teechnologies on the priority for operators, withouut compromising the quality of
device energy consumption and the respoonse delay. Unlike service (QoS) and network robustness.
r However, predicting
[9], the results showed that the timers shouuld be dynamically an accurate timing for the UEs to be awake is a very
configured and adaptive to the user’s conndition (e.g. battery challenging issue, due to thee unexpected user behavior and
restriction), and should not be statically configured by the wide variety of available servvices and applications.
serving network. However, this mightt incur additional
processing overhead at the operator side.
2) Device Traffic Management
The optimization techniques used in this approach are
seamless for the serving network. Insteead of transmitting
messages immediately after being generrated by a mobile-
side application, the mobile OS will queuue the messages to Figure 1:
1 DRX Cycle
be sent in bulk, thus ensuring periods ofo non-activity and
fewer transitions between states. Since delaying sensitive The DRX framework was implemented in the cellular
information will degrade the user experrience, the mobile networks to improve the tradde-off between QoS and power
device OS will change the traffic pattern only for the delay- consumption at the UE sidde. Moreover, LTE DRX was
tolerant applications. The authors in [11] suggested three designed to adapt quickly too different user conditions. The
optimization algorithms: piggybackingg, batching, and DRX, as described in [7], represents the period where the
TailEnder. In [12], it was shown thatt the use of one UE monitors the PDCCH; furthermore, it specifies the
optimization algorithm or a combinnation of several periods at which the operattor is able to schedule the UE
techniques can lead to an average of 40% savings in resource reserved resources.
consumption. However, this approach caannot be used with The LTE DRX (see Figuree 1) is defined by the following
real-time applications because it affects the quality of the operator-configured parameteers:
provided service.
ƒ DRX Cycle: This parametter is represented by the On
3) Device and Network Cooperation (Fasst Dormancy) Duration Period where the UE is in active mode and
The serving network cannot predict the user’s traffic monitors the PDCCH, and theen followed by possible inactive
pattern; however, the UE can provide an insight of its period. The DRX Cycle vallue is set by the operator and
application behavior. This information caan be useful at the depends on the required QoS.
network side in order to reduce the tail timers, and hence to ƒ On-Duration Timer: Thiss represents the number of
switch the UE faster into a more efficiennt power state and consecutive PDCCH sub-fram mes at the establishment of the
release network resources. This mechaniism is called “Fast DRX Cycle where the UE is i active. This timer gives the
Dormancy” (FD). FD was standardized by 3GPP in Release network the capability to tigghtly divide the time between
8 [4], where the UE notifies the networkk when there is no DRX UEs and hence to configure
c the control channel
more data to be transferred. The UE sends a Signaling resources efficiently.
Connection Release Indication (SCRI) message with the ƒ DRX-Inactivity Timer: Thiis value denotes the number of
cause set to “UE Requested PS Data Session End”. When the consecutive successfully decoded PDCCH sub-frames after
serving network receives this type of SCRI message, it which the data will be trannsferred from/to a certain UE.
releases the tail timers and relieves thhe UE to a more Moreover, it supplies the netw
work with the period to maintain
efficient power state. According to [13],, FD has not been the UE in an active mode beyyond the On-Duration timeout.
adopted by a significant number of netw work operators and
handset vendors. We are aiming in this papper to empower the UE with a
III. BACKGROUND forecasting algorithm allowinng it to decide whether it should
stay in LTE_ACTIVE or move m to LTE_IDLE in order to
The operator tightly controls the schheduling of users’ save energy. The algorithm m expects the next message to
power behavior in LTE. At the shared downlink
d interface, appear after T seconds from the
t instant the decision is made.
the user frequently checks the Physical Downlink Control If T is very large, then LTE__IDLE would be beneficial over
Channel (PDCCH) to verify whether or not the network LTE_ACTIVE, while very small s T results in the opposite
allocated the requested resources. In the case of first selection. The value of T thhat equalizes the two costs of
transmission failure, the UE informs the serving
s network by power and signaling is calledd LTE_IDLE timeout threshold,
sending a Negative Acknowledgment (NA ACK) message. The and will be determined in the next section.

195
IV. DETERMINING LTE_IDLE TIMEOUT THRESHOLD LTE_ACTIVE long DRX. 0 < DA < DI.
TMSIRA: Cost for TMSI reallocation.
Intelligently switching states between LTE_ACTIVE and
TA: Time period where UE stays in
LTE_IDLE can reduce bandwidth cost and network
LTE_ACTIVE long DRX.
congestion if tuned and optimized correctly. Otherwise, it
TI: Time period where UE stays in
can cause serious performance degradation and unnecessary
LTE_IDLE.
signaling traffic. Any fixed timeout value will result in poor
customer experience for at least one class of users. If a long C: Cost to be reduced while being in listening
timeout value is used, low-traffic users (especially ones who mode per time unit.
are not using always-on applications) will experience low FW: First Wait is the minimum drx-
battery life in addition to excessive signaling. On the other InactivityTimer, which represents the time
hand, using short timeout values will result in short battery difference between the last message and
life and excessive signaling for high-traffic users (especially the arrival of a DRX Command MAC
with applications sending periodic keep-alive messages). control element.
In this section, the LTE_IDLE timeout threshold will be
modeled. Based on this model, two timeout expressions will The left hand side of the inequality (1) accounts for the
be derived, depending on the targeted cost to be optimized, cost of transition to LTE_IDLE using a release message,
either for reducing energy consumption or reducing whereas the right hand side of (1) models the cost of staying
signaling traffic. in LTE_ACTIVE long DRX. We determine the threshold

‫ ܫʹܣ‬൅ ‫ ܣʹܫ‬൅  ܶ‫ܫܵܯ‬ோ஺  െ ‫ ܴܺܦʹܥ‬െ ‫ܥʹܴܺܦ‬

We are interested in deciding for a specific silence period using:

݄ܶ‫ ݈݀݋݄ݏ݁ݎ‬ൌ 
whether it is better to move to LTE_ACTIVE Long DRX or
ሺ‫ ܣܦ‬െ ‫ ܫܦ‬ሻ ൈ ‫ܥ‬
Ǥ
directly transit into LTE_IDLE. The overall cost for being in
LTE_IDLE state during an inactivity period (silence period)
of length T is the sum of all the following operations: If the silence period is greater than the threshold, then it is
ƒ The cost of transiting a UE to LTE_IDLE. more adequate for the UE to transit into LTE_IDLE;
ƒ The cost of staying in LTE_IDLE for T time units. otherwise, staying at LTE_ACTIVE long DRX will be less
ƒ The cost of transiting to LTE_ACTIVE. expensive. An optimal state has to be decided for each
ƒ The cost of reallocating the Temporary Mobile silence period, according to the optimization class (power or
Subscriber Identity (TMSI). signaling optimization) adopted by the operator. Based on
The overall generic cost for remaining in LTE_ACTIVE the threshold expression, two optimization models are
long DRX state during the same inactivity period T is the proposed, the first one aims to optimize the signaling
sum of all the following operations: overhead, while the second optimizes the power
ƒ The cost of transiting a UE to LTE_ACTIVE long DRX. consumption at the UE.
ƒ The cost of remaining at LTE_ACTIVE long DRX for T
time units. SIGNALING LOAD OPTIMIZATION MODEL - the signaling
ƒ The cost of transiting to LTE_ACTIVE with a DRX cost is estimated as follows:
value suitable for call setup. A2I = 1 signaling message (uplink).
I2A = 36 signaling messages, with an average delay of 188
For the transition to the LTE_IDLE state to be beneficial ms [14].
during an inactivity period T, the overall cost for staying in DRX2C = 2 signaling messages with back-off timers with
LTE_IDLE should be less than the overall cost for staying in an average delay of 13 ms [14].
LTE_ACTIVE, as shown in the following inequality: C2DRX = 2 signaling messages.
‫ ܹܨ‬൅ ‫ ܫʹܣ‬൅  ܶூ ൈ  ‫ܦ‬ூ  ൈ ‫ ܥ‬൅ ‫ ܣʹܫ‬൅  ܶ‫ܫܵܯ‬ோ஺ ൏ ‫ ܹܨ‬൅
TMSIRA = 2 signaling messages.
‫ ܴܺܦʹܥ‬൅  ‫ܦ‬஺  ൈ ܶ஺ ൈ ‫ ܥ‬൅ ‫ܥʹܴܺܦ‬
The LTE_IDLE timeout value to optimize signaling traffic
(1)

ͳ ൅ ͵͸ ൅ ʹ െ ʹ െ ʹ ͵ͷ

is therefore:

ൌ 
ሺ‫ ܣܦ‬െ ‫ ܫܦ‬ሻ ൈ ‫ܥ‬ ሺ‫ ܣܦ‬െ ‫ ܫܦ‬ሻ ൈ ‫ܥ‬
Ǥ
The parameters in (1) are defined in Table 1.

Table 1. Parameter Definitions for Cost Function Where C is the average number of paging messages
received by UE within time period T.
A2I: Cost for transiting to LTE_IDLE state POWER CONSUMPTION OPTIMIZATION MODEL - the
I2A: Cost for transiting to LTE_ACTIVE state parameters in this case are as follows:
C2DRX: Cost for transiting from LTE_ACTIVE A2I = Power consumed during sending the release
(call mode) to LTE_ACTIVE long DRX message.
DRX2C: Cost for call setup from long DRX. I2A = Power consumed during the transition from
D I: (ON-Duration/DRX cycle) in LTE_IDLE. LTE_IDLE to LTE_ACTIVE.
0 < DI < 1 DRX2C = Power consumed during the transition from
D A: (ON-Duration/DRX cycle) in LTE_ACTIVE to call setup mode.

196
 C2DRX = Power consumed during the transition from call decision algorithm, the factor SFa was added to the model.
setup mode to LTE_ACTIVE long DRX. When SFa tends to 1 (I am 100% confident that the duration
TMSIRA = 0 since TMSI generation is implemented in till next packet silence period is greater than timeout

‫ ܫʹܣ‬൅ ‫ ܣʹܫ‬െ ‫ ܴܺܦʹܥ‬െ ‫ܥʹܴܺܦ‬

MME and thus does not require computation at the UE side. threshold), the transition requirements into LTE_IDLE are

݄ܶ‫݈݀݋݄ݏ݁ݎ‬ሺ‫ݐ݌݋‬Ǥ ‫ݎ݁ݓ݋݌‬ሻ ൌ 
satisfied, and hence the UE will most probably transit into
ሺ‫ ܣܦ‬െ ‫ ܫܦ‬ሻ ൈ ‫ܥ‬
Ǥ
LTE_IDLE. On the other hand, when SFa tends to 0, (I have
no confidence that the duration till next packet silence period
where C is the number of power units consumed within one
is greater than timeout threshold), the decision criteria to
unit of time while being in listening mode.
transit into LTE_IDLE becomes harder to be satisfied, which
means the UE will most probably decide to transit into
V. DECISION ALGORITHM LTE_ACTIVE long DRX. It is left for the device/OS
In the previous section, we derived the threshold value for manufacturer or user to tune the SFa value.
LTE_IDLE timeout. We also proposed two LTE_IDLE Let X be a random variable representing the number of
timeout threshold models; the first optimizes the UE’s power messages generated by any of the above three traffic groups
consumption while the other optimizes the signaling in duration equal to LTE_IDLE timeout threshold. Pr(X=0)
overhead. The challenging part is to estimate, ahead of time, is the probability that no incoming or outgoing traffic has
the UE’s silence duration. In other words, the threshold occurred within LTE_IDLE timeout threshold.
value is useless unless the next packet exchange can be If Pr(X)  SFa, the UE should transit to the LTE_IDLE state.
predicted. The UE applies the decision algorithm to either In other words, if it is most probable that no traffic will be
switch into LTE_IDLE if the silence duration is greater than received within the coming 35C/(DA - DI) units of time, the
the threshold, or to stay in LTE_ACTIVE long DRX, mobile device should switch to LTE_IDLE state.
otherwise. AP and LS are independent positive values whereby:

‫‘••‹‘̱ܲܣ‬ሺ‫ܲܣ‬஺௩ ሻ
The LTE_IDLE timeout should satisfy the following

‫‘••‹‘̱ܵܮ‬ሺ‫ܵܮ‬஺௩ ሻ
constraints:

‫ ܲܣ‬൅ ‫‘••‹‘̱ܵܮ‬ሺ‫ܲܣ‬஺௩ ൅  ‫ܵܮ‬஺௩ ሻ

ƒ It tends to be minimal if the UE is in power-saving mode.
ƒ If the battery is nearly discharged, we need to keep the UE
in LTE_IDLE to extend the battery life as much as
possible. The time interval of consecutive keep-alive messages
ƒ For optimal battery performance, the UE OS should block corresponding to application i (h[i]) is periodic, which means
all connections other than the ones initiated by the user. that the jth keep-alive message of the ith application (h[i][j])
The overall user traffic is divided into three groups: is known in advance.

‫ܲܣ‬ԡ‫ܵܮ‬ԡ݄ሾ݅ሿሾ݆ሿ݀‫݁ݎ݋݂ܾ݁ݎݑܿܿ݋ݐ݋݊ݏ݁݋‬
1) The incoming and outgoing traffic of always-online
”ሺܺሻ ൌ ” ൬ ൰
‫݈݀݋݄ݏ݁ݎ݄ݐݐݑ݋݁݉݅ݐ ܧܮܦܫܧܶܮ‬
applications without their corresponding keep-alive
messages. The number of session arrivals per 35C/(DA -

Ͳ ǥ ǥ ǥ ǥ Ǥ Ǥǡ ‫ݏݎݑܿܿ݋݁ݒ݈݅ܽ݌݂݁݁݇ܫ‬
DI) units of time related to always-online applications is

”ሺܺሻ ൌ ൝ ‫ܧܮܦܫ̴ܧܶܮ݄݊݅ݐ݅ݓ‬
represented by a random variable AP, which follows a

݁ ି஺௉ಲೡି௅ௌಲೡ ǡ ܱ‫݁ݏ݅ݓݎ݄݁ݐ‬
Poisson distribution of average rate APAV, which is a UE
calculated parameter that shows the average of UE's

݁ ି஺௉ಲೡ ି௅ௌಲೡ ൒ ܵ‫ܽܨ‬

incoming and outgoing traffic in number of messages.
2) The incoming and outgoing traffic of legacy services
(SMS, calls, browsing). The number of arriving legacy
calls per 35C/(DA - DI) is represented by a random To simplify the expression, we take the logs of both sides
variable LS, that follows a Poisson distribution of average of the inequality to get:

‫ܲܣ‬஺௩ ൅  ‫ܵܮ‬஺௩ ൑ Žሺܵ‫ܽܨ‬ሻ

of LSAv.
3) Keep-alive messages, each related to an installed always-
online application. The time interval of consecutive keep-
alive messages corresponding to application i is Summing up, keep-alive messages are periodic and can be
represented by a random variable h[i], where h[i][j] is the predicted successfully as we will show in Section 6. The
jth keep-alive message of the ith application. signaling optimization formula showed that the LTE_IDLE
The values APAv and LSAv can be calculated from the timeout threshold is equal to 35C/(DA - DI). By identifying
traffic averages within designated time frames (for example, DA and DI values, the threshold will be specified such that
hour, day, or week) or by comparing similar periods across the signaling costs of staying in LTE_ACTIVE state or
time (for example, the same weekday for several weeks, or a switching to LTE_IDLE state are equal. Considering a
specific day of the month for several months). Such confidence level of 50% (SFa = 0.5, ln(SFa)=0.6931), then
implementation is left for the operator to decide. the final decision algorithm will be as follows:
Since we are not able to predict exactly the duration of the
silence period, we will try to estimate it with high
confidence. To specify the confidence value of the applied

197
 OR (‫ܲܣ‬஺௩ ൅  ‫ܵܮ‬஺௩ < 0.6931)) then
if ((Device is in power save mode)

Go to LTE_IDLE
else
Go to LTE_ACTIVE long DRX

VI. APPLICATION PROFILLING Figure 3: WhatsApp Keep-alive Messages

In the previous section, we have connsidered that keep-

alive messages follow a deterministic law w. To confirm the
validity of our assumption, packet sniffingg experiments were
implemented and the results are provided ini this section.
The experimental setup, shown in Figurre 2, consists of an
Android-based smartphone running alll the applications
under investigation. The packet sniffer sooftware, Wireshark,
Figure 4: IM+ Keeep-alive Messages
was installed on a laptop in order to cappture all the uplink
and downlink traffic on the smartphonne, and finally an
We investigated two always-online push applications:
access point was used to connect the smartphone to the
WhatsApp and IM+. Ressults showed that WhatsApp
Internet. It is important to note that since
s we are only
exchanges aggregated keeep-alive and status updates
interested in the traffic generated by the application,
a and the
messages with the server evvery three minutes as shown in
signaling traffic with the Wi-Fi radio inteerface does not get
Figure 3. This exchange requuires 20 packets with a total of
forwarded to the core network, the data packets generated
2500 bytes. On the other hannd, IM+ sends 10 packets with a
are the same whether on a Wi-Fi coonnection or on a
total of 1000 bytes for the keep-alive
k messages every 150
2.5/3G/4G connection. Therefore the results obtained from
seconds as shown in Figure 4, and sends 50 packets with a
this study can be adopted for the backkbone of a packet-
total of 10,000 bytes for the status updates every 600
switched core network.
seconds. Both applications shhowed that keep-alive messages
follow a deterministic pattern;
p therefore their next
transmission time can be easily predicted.
VII. KEEP-ALIVE TRAFFIIC PATTERN AUTO-LEARNING
ALGO
ORITHM
Applications start sending their corresponding keep-alive
messages after a certain tim me in their loading phase. Since
applications differ in size andd require processing for loading,
Figure 2: Experimental Seetup keep-alive messages of diff fferent applications differ with
random phase shift, and varrying frequencies. The UE OS
It is important to emphasize that sm martphone always- does not currently need cooperation from installed
online applications have a different behhavior than regular applications to get inform med about the corresponding
voice applications regarding their networkk communications. periodicity and phase shift of keep-alive messages. We
In fact, the study in [15] has categorizedd these applications present next a technique for the OS to collect and infer the
into three categories, according to the characteristics of periodicity and phase shift of different transmitted keep-
maintaining their connectivity to the serveer. The pull/polling alive messages. We proposee in this section an algorithm
category includes applications whereby the initial request allowing the UE to learn andd start estimating the pattern of
originates from the user and immediatelyy processed by the different keep-alive messagess.
server. Such applications include Faceboook, Twitter, and the Two methods are proposeed to auto-learn the keep-alive
like. The second category, the long pollingg, is a variation the traffic pattern: one is based on o sequences, while the other is
pull/polling. After a connection requestt to the server is based on Fourier series.
initiated by the user, the connection iss kept active until Numeric sequences metthod: Suppose we have N
updates from the server are available or until
u the connection applications, each of which generates keep-alive messages
times out. Finally, the third category is the push category, with period Ti phase shift i. The timestamp h[i][j] of the jth
where the user sends keep-alive messages to the server and keep-alive message generatedd by the ith application is:
݄ሾ݅ሿሾ݆ሿ ൌ  ܶ௜ ൈ ݆ ൅  ߠ௜ j ‫א‬IN+ ……..(2)
the latter pushes updates and notificaations to the user
whenever the need arises. Such appplications include
WhatsApp and IM+. In order to prove that t the keep-alive An application i1 of h[i1][jj1] = j1 × Ti1 + i1 is hidden by
application i2 of h[i2][j2] = j2 × Ti2 + i2
ܶ௜ଵ െ  ܶ௜ଶ ܶ௜ଵ
messages can be predicted, we are only intterested in the push

݂݅ ‫ܰܫ א‬
ܰ ା ƬƬ  ‫כܰܫ א‬ା
applications, since the other two categorries do not use the
ܶ௜ଶ ܶ௜ଶ
keep-alive technique to maintain the serveer connection.

198
 ାஶ
ͳ ௜ଶగ௞ሺ௧ିఏ ሻȀ் ͳ ௜ଶగ௞ሺ௧ିఏ ሻȀ்
Ɂ୘ ሺ‫ݐ‬ሻ௧௢௧௔௟ ൌ ෍ ൬ ݁ భ ൅ ݁ మ൰
The applications are organized in this representation using
ܶଵ ܶଶ
the following criteria: T1 < T2 <…< TN-1 < TN భ మ

The convergence period is CP = 2(TN + N) ିஶ

݇ ቀ ቁ ൌ ܽǢ ܽ ‫ א‬
The traffic is received when:
௧ିఏ೔
Consider that within CP, keep-alive messages/pulses have

‫{א‬P1,P2,….,PN}. Let Z be [N/2], so we have Z groups:

occurred and the time of ith keep-alive message/pulse is Pi
்೔
௔
ܲଶ െ ܲଵ Ͳ Ͳ Ͳ Ͳ ௞
ܲଷ െ ܲଵ ܲଷ െ ܲଶ Ͳ Ͳ Ͳ ‫ ݐ‬ൌ ݆ ൈ ܶ௜ ൅  Ʌ௜
Assume jൌ then:

‫ۇ‬ ‫ۊ‬
‫ڭ‬ ܲସ െ ܲଶ ‫Ͳ ڰ‬ Ͳ
‫ۈ‬ ‫ۋ‬
ܼൌ‫ڭ ۈ‬ ‫ڰ‬ Ͳ
We have achieved the same equation as that using the
‫ۋ‬
‫ڭ ۈ‬ ܲ௭ାଵ െ ܲ௭ ‫ۋ‬
numeric sequences method.

‫ڭ‬ ‫ڭ‬
‫ܲۉ‬ே െ ܲଵ ܲே െ ܲଶ ǥ ǥ ܲே െ ܲ௭ ‫ی‬
VIII. EXPERIMENTAL VALUES AND ANALYSIS
In this section, we present simulation results for the
behavior of 150 UEs, where the traffic generated by
Each column is studied separately. Consider the first applications and legacy calls follow Poisson distributions.
group: {P2 – P1, P3 – P1, ..., PN – P1}. The period of group k The simulated scenarios are categorized into 5 classes
ܲ௡ െ  ܲ௞
is dk:
݂݅‫ ݅׊‬൏ ‫ ݆׌‬൑ ݊Ȁ݀௞  ൈ ݅ ൌ ܲ௝ି ܲ௞
according to the user’s usage regarding legacy calls and

݀௞
application traffic. We will refer to the decision algorithm

LTE_IDLE timeout threshold ሺ୍ ሻ୮୭୵ୣ୰୭୮୲ as “dynamic

using the proposed dynamic LTE_IDLE timeout based on
Then hk(j) = dk × j + Pk
timeout”. We will also refer to fixed LTE_IDLE timeouts (as
P1 is generated by h[N][j] and its first captured pulse (j=0),
currently implemented in 3G and LTE networks) as “fixed
then P1 = N. In the first column, N is subtracted, thus other
timeouts”. The performance of dynamic timeout will be
period (ܶ௡ ൈ ݆). In the above expression, we are looking for
pulses from the same application will occur as multiple of its
compared with different fixed timeout values to measure the
terms that are multiples of each other (ܶே ሻ.
resulting enhancements.
The studied scenarios are classified according to the user
The second method we used relies on the Fourier series:
behavior as follows:
The keep-alive traffic of an application is a periodic burst of
ƒ Class 1: low to moderate legacy traffic with no installed
shown in Figure 5.a, whereby application ݅ can be
bytes so it can be represented as a square wave signal as
always-online applications.
represented by 3 parameters ሺߠ௜ ǡ ܶ௜ ǡ ߬௜ ሻ, where ܶ is the
ƒ Class 2: moderate to high legacy traffic with no installed
signal period, ߬ is the burst duration, and ߠ is the initial
always-online applications.
ƒ Class 3: moderate legacy traffic with low application
phase shift of the first keep-alive message. traffic is installed.
The burst duration is constant because the keep-alive size ƒ Class 4: moderate legacy traffic with moderate application
is smaller than a Resource Block. Furthermore, since we are traffic.
only concerned with the pulse starting time, each keep-alive ƒ Class 5: moderate legacy traffic with high application
message can be modeled as a Dirac pulse (Figure 5.b). traffic.
According to [10], the power consumption in PCH state is
two times the one in IDLE state in 3G networks. We will
apply the same ratio between LTE_ACTIVE long DRX and
LTE_IDLE. We shall consider:
DI × C = 1 power unit
DA× C = 2 power units
A2I, I2A, C2DRX and DRX2C are considered proportional
to the number of transmitted/received messages, thus:
Figure 5: Keep-Alive as a Dirac wave A2I: 1 (relative unit)
I2A: 36

ͳ ஶ ௜ଶగ௞௧Ȁ்
The Fourier Series of Dirac wave is given by: C2DRX: 2
Ɂ୘ ሺ‫ݐ‬ሻ ൌ ෍ ݁ 
ܶ
DRX2C: 2
ି௪ Thus, TI > (1+36-2-2)/(2-1) => TI > 33 s.

ͳ ஶ ௜ଶగ௞ሺ௧ିఏభ ሻȀ்భ 
Assuming a smartphone is running two applications:

Ɂ୘ଵ ሺ‫ݐ‬ሻ ൌ ෍ ݁ 

ܶଵ ିஶ
ͳ ஶ ௜ଶగ௞ሺ௧ିఏమ ሻȀ்మ 
Ɂ୘మ ሺ‫ݐ‬ሻ ൌ ෍ ݁ 
ܶଶ ିஶ
The cumulative traffic will be:

199
 Fixed Vs Dynamic Timeout (mSec)
5,000 15,000 25,000 35,000 45,000 55,000 65,000 75,000 85,000 1E+20
0
Power Consumption Overhead (%)

-5

-10

-15

-20

-25

-30

-35

-40

-45

Figure 7:Average Power

Class 1 Consumption:
Class 2 Class 3 Fixed
Class 4 VsClassDynamic
5 Timeout
Figure 6: Average Power Consumption Comparison For User’s Classes

In LTE_ACTIVE long DRX, the UE sends message updates represented experimentally similar to class 1 but with double
on each cell handover, while in LTE_IDLE, the UE sends the call establishment rate. Short timeout values are suitable
message updates on each location area handover [1]. The for low traffic users, while long timeout values are suitable
number of sent messages depends on the network scheme for those with very high traffic. In case of moderate traffic
(number of cells per location area, location area scheme: users, the performance of different timeout values is nearly
overlapping, collocated...). We haven’t considered this similar. It can be noticed for class 5 users (very high traffic)
aspect in our study, thus we will only show the resultant that higher timeout values have better overall performance.
expression to be studied later in greater depth. Similarly, short timeout values have similar overall
Figure 6 shows the results of our proposed decision performance to long values. In addition, applying our
algorithm for the five user behavior classes mentioned algorithm on class-2 users will achieve 13 to 20% savings in
above. The output is compared with fixed LTE_IDLE energy consumption.
timeout values. The x-axis represents the fixed timeout
compared with our dynamic value, while the y-axis Class 3: In this class, users are generating moderate
represents the resulting power consumption overhead. The legacy traffic with low application traffic. This class is
"1E+20" value is an infinite value for LTE_IDLE timeout, similar to most of the current smartphone users. Each user in
and represents the case when the UE is always in the this class is represented experimentally similar to class 1 for
LTE_ACTIVE state. the legacy traffic, in addition to a random combination of
applications each having one of the following keep-alive
Class 1: In this class, only legacy calls (voice calls, SMS, intervals: 150s similar to Yahoo IM, 250s similar to Viber,
browsing) are considered where the users don’t have always- and 300s similar to WhatsApp. The number of calls per
online applications installed on their mobile phones. The application is random with an average of 5 calls per hour,
users generate traffic at low to moderate rates. Each user in and the call duration of an application is also random. The
this class is represented experimentally by the establishment performance comparison for class-3 users achieves similar
of legacy call establishments following a Poisson results to class-1 users and this is due to low/moderate traffic
distribution with a rate of 1 call per hour, call duration rates. It is shown in Figure 6 that for class-3 users, our
following a normal distribution of average 120s and standard algorithm is able to achieve 3 to 25% savings in energy
deviation 30s. In low traffic scenarios, staying long in consumption.
LTE_ACTIVE state will result in more power consumption,
Class 4: In this class, users generate moderate legacy
thus long timeout values degrade performance. As shown in
traffic with moderate application traffic. Each user is
Figure 6, the short timeout values have better overall
represented experimentally similar to class 3 but with double
performance than long values. It can also be shown that for
the application call establishment rate. The performance
class 1 users, our algorithm is able to achieve 12 to 30%
comparison for class-4 users achieves similar results to
savings in energy consumption.
class-2 users and this is due to low/moderate traffic rates. It
Class 2: This scenario is similar to class 1 except that the is shown in Figure 6 that for class-4 users our algorithm is
users generate moderate to high traffic rates. Each user is able to achieve 5 to 42% savings in energy consumption.

200
 Class 5: In this class, users generate moderate legacy [4] 3GPP TS 25.331 V11.3.0. Radio Resource Control (RRC);
traffic with high application traffic. Each user is represented Protocol specification. (2012-09).
experimentally similar to class 4 but with triple the [5] 3GPP TS 36.304 V11.1.0. User Equipment (UE) procedures in
application call establishment rate. Short timeout values are idle mode. (2012-09).
suitable for low traffic users, while longer timeout values are [6] 3GPP TS 36.321 V11.0.0. Medium Access Control (MAC)
more appropriate for those with very high traffic. This can be protocol specification. (2012-09).
seen for class-5 in Figure 6 where high long fixed timeouts [7] 3GPP TS 36.331 V11.1.0. Radio Resource Control (RRC);
have better performance than short fixed timeouts. It is also Protocol specification (2012-09).
shown that for class-5 users our algorithm is able to achieve
[8] Mohan, Siddharth, Rohit Kapoor, and Bibhu Mohanty.
7 to 30% savings in energy consumption. "Latency in HSPA Data Networks." QUALCOMM. Available:
http://www.Qualcomm.com/documents/files/latency-in-hspa-
Overall, for the different traffic classes, our proposed data-networks.pdf, [27 January 2011] (2011).
algorithm is able to decrease energy by 20% on average. [9] M. Chuah, W. Luo, and X. Zhang. Impacts of Inactivity Timer
Since every network has users belonging to a combination of Values on UMTS System Capacity. In Wireless
the above-mentioned classes, whichever chosen fixed Communications and Networking Conference, 2002.
timeout value will definitely result in poorer customer [10] Chi-Chen Lee, Jul-Hung Yeh, and Jyh-Cheng Chen, “Impact
satisfaction for an important portion of the operator’s users. of inactivity timer on energy consumption in WCDMA and
cdma2000”, in Proc. Wireless Telecommunications
IX. CONCLUSIONS AND FUTURE WORK
Symposium, 2004.
Optimal network performance is achieved when both the [11] Periodic Transfers in Mobile Applications: Network-wide
network-side and the user-side efforts and data are Origin, Impact, and Optimization
synergized. Fixed timeout transition between different states,
[12] N. Balasubramanian, A. Balasubramanian, and A.
as currently implemented in 3G and LTE networks, is just a Venkataramani, “Energy Consumption in Mobile Phones: A
network-side effort, with no support from UEs. We have Measurement Study and Implications for Network
proposed and mathematically derived in this paper a Applications”. In IMC 2009.
collaborative algorithm to decide when it is optimal for a UE
[13] F. Qian, Z. Wang, A. Gerber, Z. M. Mao, S. Sen, and O.
to transit between states. This decision is based on Spatscheck, “TOP: Tail optimization protocol for cellular radio
probabilistic estimation using a traffic pattern model resource allocation,” in Proc. IEEE ICNP, 2010
generated from an auto-learning algorithm running on the
[14] S. Mohan, R. Kapoor, and B. Mohanty, “Latency in HSPA
UE side. This intelligent decision algorithm can be tuned to Data Networks ”, white paper for Qualcomm, 2011
meet different optimization requirements. Battery
consumption and signaling overhead are the investigated [15] HUAWIE Smart Lab. (2010). ”Behavior Analysis of
Smartphone”. HUAWEI TECHNOLOGIES CO. Ltd.,
requirements in this paper, but the proposed algorithm can
Shenzhen, China.
be tuned based on other preferences as well. After
experimentally evaluating the proposed algorithm, a 10 to
30% energy savings, corresponding to a decrease of 10 to
30% in battery consumption, was achieved.

ACKNOWLEDGMENT
This research was funded by TELUS Corporation, Canada.

REFERENCES
[1] 3GPP TS 23.122 V11.3.0. Non-Access-Stratum (NAS)
functions related to Mobile Station (MS) in idle mode. (2012-
09).
[2] 3GPP TS 23.401 V8.10.0. General Packet Radio Service
(GPRS) enhancements for Evolved Universal Terrestrial Radio
Access Network (E-UTRAN) access. (2010-06).
[3] 3GPP TR 23.882 V8.0.0. 3GPP System Architecture
Evolution: Report on Technical Options and Conclusions.
(2008-09).

201