Performance of CS Fallback for Long Term Evolution Mobile Network

Performance of CS Fallback for Long Term Evolution Mobile Network

Ren-Huang Liou, and Yi-Bing Lin, Fellow, IEEE Department of Computer Science

National Chiao Tung University {rhliou, liny}@cs.nctu.edu.tw

Yingrong Coral Sung

Center for Information and Communications Technology National Chiao Tung University

ycsung@cs.nctu.edu.tw Pang-Cheng Liu

Department of Computer Science National Chiao Tung University

pcliu@cs.nctu.edu.tw Christian Wietfeld

Communication Networks Institute TU Dortmund University Christian.Wietfeld@tu-dortmund.de

Abstract

When a mobile operator migrates its network from the 3rd Generation (3G) system to Long Term Evolution (LTE), both 3G and LTE will co-exist for a period of time. Since the 3G Circuit-Switched (CS) voice mechanism is more mature and available than that for LTE Voice over Internet Protocol (VoIP), the operator may consider CS fallback as a solution to provide reliable voice calls. According to the 3rd Generation Partnership Project (3GPP) CS fallback procedure, when a mobile user in the LTE network has an incoming or an outgoing call, the User Equipment (UE) falls back from LTE to Universal Mobile Telecommunications System (UMTS). When the call is complete and released, the UE immediately returns to LTE. If the next activity for the UE is another voice call, immediately switching from UMTS

to LTE may not be efficient. In this case, the UE has to perform another CS fallback. To resolve this issue, we suggest delaying the returns to avoid unnecessary CS fallbacks, which is called delayed-return (DR). Based on the measurements from the real UMTS and LTE networks, we develop analytic model to investigate the performance of the CS fallback with DR. The study indicates that the DR scheme can effectively reduce the CS fallback costs up to 60%.

Index Terms: CS fallback, delayed-return scheme (DR), long term evolution (LTE)

1 Introduction

The 3rd Generation Partnership Project (3GPP) Long Term Evolution (LTE) [1] defines an all IP network architecture that evolves from the Universal Mobile Telecommunications System (UMTS) [2] to provide high-speed data services. When a mobile operator migrates its network from the 3rd Generation (3G) system to LTE, both 3G and LTE will co-exist for a period of time, and initially, the 3G coverage is more complete than the LTE coverage.

Since LTE is a Packet-Switched (PS) network, the Circuit-Switched (CS) voice service [3] is not supported as 3G does. Therefore the LTE voice service is offered through the Voice over Internet Protocol (VoIP) technology. Since the 3G CS voice mechanism is more mature and available than that for LTE VoIP, in the deployment of LTE, many operators consider CS fallback [4] as the solution to provide reliable voice calls. The CS fallback technique switches the User Equipment (UE; the mobile phone) from the LTE network to the 3G legacy system when a voice call is attempted. In order to use the CS domain in the UMTS system when the UE resides in LTE, the LTE network needs to register the UE with both the LTE and the UMTS CS domain and delivers the CS paging message from UMTS to the UE.

Figure 1 illustrates a simplified architecture of UMTS and Evolved Packet System (EPS) for LTE. This architecture includes two parts: the UMTS network and the LTE network.

MSC/VLR

SGSN

MME

CS Network

S-GW

P-GW External

Data Network HSS

11 1

6

9

5

7 4

10 UMTS Network

LTE Network a

b UE

NodeB

RNC 12

eNodeB 3 2

8

Figure 1: The EPS and UMTS Architecture for CS Fallback (dashed lines: signaling; solid lines: signaling/data)

A UE (Figure 1 (1)) accesses UMTS and LTE services through the radio interfaces. In the UMTS network (Figure 1 (a)), the UMTS Terrestrial Radio Access Network (UTRAN) consists of NodeBs (Figure 1 (2)) and Radio Network Controllers (RNCs; Figure 1 (3)).

A NodeB provides Wideband Code Division Multiple Access (WCDMA) radio connectivity between the UE and the corresponding RNC. The RNC connects to the UMTS core network.

This core network is partitioned into the CS and the PS domains. The CS domain includes Mobile Switching Centers (MSCs) and Visitor Location Registers (VLRs; Figure 1 (4)). An MSC is responsible for call control and connection between the UE and the external CS Network (Figure 1 (5)). A VLR is responsible for the mobility activities of the MSC. The PS domain consists of Serving GPRS Support Nodes (SGSNs; Figure 1 (6)) which provide the mobility and session services to the UEs. The Home Subscriber Server (HSS; Figure 1 (7)) is the master database containing all user-related subscription information, which supports mobility management of mobile users. In the LTE network (Figure 1 (b)), the Evolved UMTS Terrestrial Radio Access Network (E-UTRAN) consists of evolved NodeBs (eNodeBs; Figure 1 (8)) to offer LTE radio connectivity to the UE. The E-UTRAN connects to the LTE core network that includes the following components. A Mobility Management

Entity (MME; Figure 1 (9)) interacts with the HSS to offer mobility management and session control. The Serving Gateway (S-GW; Figure 1 (10)) is responsible for routing data packets and is an anchor of the user plane data for intra- and inter-system handovers. The Packet Data Network Gateway (P-GW; Figure 1 (11)) provides the connectivity to the External Data Network (Figure 1 (12)) and the per-user based packet filtering. In the UMTS PS domain, the SGSNs connect to the External Data Network through the S-GW and the P- GW. According to the 3GPP CS fallback procedure [4], when a mobile user in the LTE network has an incoming or an outgoing call, the UE falls back from LTE to UMTS. When the call is complete and released, the UE immediately returns to LTE. If the next activity for the UE is another voice call, immediately switching from UMTS to LTE may not be efficient.

In this case, the UE has to perform another CS fallback. To resolve this issue, we suggest delaying the returns to avoid unnecessary CS fallbacks. This paper is organized as follows.

Section 2 describes the CS fallback procedures, the existing Immediate-Return (IR) scheme and the proposed Delayed-Return (DR) scheme for returning to LTE. Section 3 proposes an analytic model for the IR and the DR schemes. Section 4 studies the performance of IR and DR by numerical examples, and conclusions are given in Section 5.

2 3GPP CS Fallback Procedures

This section describes the CS fallback procedures defined in the 3GPP, including call setup and call release with IR. We also report the measured processing times for the proce- dures collected in live 3G and LTE networks in [5, 6, 7]. Then we introduce the DR scheme, including call release and data session setup.

13. RRC Connection Setup Complete 1. Extended Service Request

4. RRC connection release with Redirection to UTRAN 5. UE Context Release Request

14. CM Service Request

15. Call establishment procedure

6. Release Access Bearers Request 7. Release Access Bearers Response 8. UE Context Release Command 9. UE Context Release Complete

11. RRC Connection Request 12. RRC Connection Setup

2. UE Context Modification Request 3. UE Context Modification Response

10.Tune to UMTS &

Camp on cell

UE NodeB/RNC eNodeB SGSN MME S-GW MSC

Figure 2: Call Setup with CS Fallback

2.1 LTE Call Setup with CS Fallback

Figure 2 illustrates the CS fallback message flow when a UE makes a call in the LTE network. The following steps are executed:

Step 1. The UE sends the Extended Service Request message to the MME to initiate the CS fallback procedure.

Steps 2 and 3. The MME exchanges the UE Context Modification Request and Response

message pair with the eNodeB to indicate that the UE should fall back to the UTRAN.

Note that Steps 1-3 take about 0.3 seconds [5].

Step 4. The eNodeB sends the UE the Radio Resource Control (RRC) Connection Release with Redirection to UTRAN message to indicate that it may follow the cell identity and System Information to attach to the corresponding UTRAN cell.

Steps 5-9. Parallel to Step 4, the eNodeB sends the UE Context Release Request message to the MME to release the bearers between the eNodeB and the S-GW. Steps 4-9 take about 0.2 seconds [5].

Step 10. After Step 4, the UE tunes the radio to UMTS, and camps on the NodeB according to the System Information in the RRC Connection Release with Redirection to UTRAN message. Step 10 takes about 2.3 seconds for 3GPP R8 and 0.3 seconds for 3GPP R9 [5].

Steps 11-13. The UE exchanges with the NodeB the RRC Connection Request and Setup message pair to establish the radio connection. Then the UE sends the NodeB the RRC Connection Setup Complete message to acknowledge the RRC connection establishment.

Steps 11-13 take about 0.3 seconds [5].

Steps 14 and 15. The UE sends the Call Management (CM) Service Request message to initiate the CS call establishment procedure. The UE includes the Circuit-Switched Mobile Originated (CSMO) flag to indicate that it is a CS fallback call. The CS call establishment at Step 15 follows the 3GPP standard, and the details can be found in [8]. Steps 14 and 15 take about 3.5 seconds [5].

Note that if the UE in LTE is engaged in a data session when a call arrives, then the PS connection (for the data session) is also switched to UMTS in the call setup of the CS fallback procedure. Details of PS connection switching can be found in [1].

(End of CSFB flag)

(Redirection Info)

5. Iu Release Complete 2. Iu Release Command

4. Radio Bearer Release Complete

7. RRC Connection Release Complete 6. RRC Connection Release

NodeB/RNC

UE MSC

1. 3GPP call release procedure

3. Radio Bearer Release

8. Turn to LTE and camp on cell

9. 3GPP UMTS to LTE PS handover Without

data session

With data session

Figure 3: Call Release with IR

2.2 Call Release with Immediate-Return

Figure 3 illustrates the call release procedure with Immediate-Return (IR). After a voice call is released, if no UMTS data session is in progress, the UTRAN moves the UE to the LTE network immediately with the following steps:

Step 1. The standard 3GPP call release procedure is executed [8].

Steps 2-5. The MSC sends the UTRAN the Iu Release Command message to release the bearer between the MSC and the RNC. This message contains the End of CS Fallback (CSFB) flag to indicate that the call which was released is a CS fallback call. Then

the NodeB sends the UE the Radio Bearer Release message to release the radio bearer between the NodeB and the UE.

Steps 6-8. According to the End of CSFB flag, the NodeB knows that the UE is LTE capable. The NodeB sends the UE the RRC Connection Release with Redirection Info message to release the radio connection between the NodeB and the UE. Then the UE switches to the LTE network according the redirection information in the message.

If the UE is engaged in a data session when the voice call is released, then Steps 6-8 are replaced by the standard 3GPP UMTS to LTE PS handover (Step 9) [1], and the data session is moved to the LTE network.

2.3 Call Release with Delayed-Return

When a voice call is released, if the UE is engaged in a data session, then it is switched back to LTE as shown in Figure 3 (Steps 1-5 and 9). If the UE is not engaged in a data session, then it does not need to return to LTE immediately. Figure 4 illustrates the call release procedure with Delayed-Return (DR). The UE releases the radio connection and stays in UMTS in the idle mode.

Steps 1-5 of the message flow in Figure 4 is the same as the call release procedure with IR. At Step 6, the NodeB sends the RRC Connection Release without Redirection Info message.

Because this message does not contain the optional Redirection Info, the UE will not switch to LTE. This message instructs the UE to release the RRC connection, stay in UMTS, and change its status to the idle mode. Compared with the CS fallback with IR, Steps 8 and 9 in Figure 3 are saved in the CS fallback with DR.

(End of CSFB flag)

5. Iu Release Complete 2. Iu Release Command

4. Radio Bearer Release Complete

7. RRC Connection Release Complete 6. RRC Connection Release

NodeB/RNC

UE MSC

1. 3GPP call release procedure

3. Radio Bearer Release

(without Redirection Info)

Figure 4: Call Release with DR

2.4 Data Session Setup in UMTS with Delayed-Return

Suppose that DR is applied, and the UE does not return to LTE after a voice call (i.e., there is no data session in progress when the voice call is released). If the next event to the UE is a data session arrival, then it will receive the PS paging message from the UMTS NodeB. The UE is switched to LTE to establish the PS connection. The detailed steps are described as follows (see Figure 5):

Step 1. The UE executes the 3GPP Inter-Radio Access Technology (RAT) cell reselection procedure from UTRAN [9] to perform the measurement process, and then selects a LTE cell.

Steps 2-4. The UE exchanges with the eNodeB the RRC Connection Request and Setup message pair to establish the radio connection. Then the UE sends the eNodeB the RRC Connection Setup Complete message to acknowledge the RRC connection establishment

7. RRC Connection Reconfiguration

8. RRC Connection Reconfiguration Complete 1. Inter-RAT

cell reselection from UTRAN

5. Service Request

6. Initial Context Setup Request

9. Initial Context Setup Complete

10. Modify Bearer Request 11. Modify Bearer Request 12. Modify Bearer Response 13. Modify Bearer Response

UE NodeB/RNC eNodeB SGSN MME S-GW P-GW

2. RRC Connection Request 3. RRC Connection Setup

4. RRC Connection Setup Complete

Figure 5: Date Session Setup in UMTS with DR procedure.

Step 5. The UE sends the MME the Service Request message to initial the establishment of the bearer for the PS connection.

Steps 6-9. The MME exchanges with the eNodeB the Initial Context Setup Request and Complete message to establish the radio bearer between the UE and the eNodeB. The Request message also contains the bearer information between the eNodeB and the S- GW. The eNodeB sends the UE the RRC Connection Reconfiguration message to modify the bearer information of the radio connection. Then the UE sends the eNodeB the

Previous Call Arrival

t_s

time t_c

Previous Call Release

τ_p t_a

Next Call Arrival

t_c

Next Call Release

Previous Session

Arrival

Previous Session Release

t_s Next Session Arrival

Next Session Release t_p

t₁

t₂

t₃ t₄

t₅ t₆

t₇ t₈

Figure 6: Timing Diagram for Voice Call and Data Session Arrivals

RRC Connection Reconfiguration Complete message to acknowledge the radio bearer reconfiguration.

Steps 10-13. The MME sends the S-GW the Modify Bearer Request message to establish the bearer between the eNodeB and the S-GW and the bearer between the S-GW and P-GW.

We note that for data session setup, the CS fallback with DR does not incur extra overhead over IR from the network viewpoint. Specifically, Steps 10-13 in Figure 2 are executed by IR, which are the same as Steps 1-4 in Figure 5. IR also executes Steps 5-13 in Figure 5 when a data session arrives.

3 Analytic Model

This section proposes an analytic model to study the performance improvement of the DR scheme over the IR scheme. Specifically, we derive the probability p that when a voice call arrives, the UE can be connected at UMTS without CS fallback due to DR. Figure 6 illustrates a timing diagram for voice call arrivals (at t2 and t5) and data session arrivals (at t1 and t6). Let tc = t4− t² (also t7 − t⁵) be a voice call holding time. Let the inter-call arrival time ta= t5−t⁴ be a random variable with the density function fa(·), the distribution

function Fa(·), the variance V^aand the Laplace transform f_a^∗(s). Let the session holding time ts = t3 − t¹ (also t8 − t⁶) be a random variable with the mean 1/µ, and the inter-session arrival time tp = t₆− t3 be a random variable with the density function fp(·), the variance Vp and the Laplace transform f_p^∗(s). Suppose that the call release event at t₄ is a random observer of the period [t3, t6]. From the residual life theorem [10], the interval τp = t6 − t⁴ is the residual life of tp with the density function rp(·), the distribution function R^p(·), and the Laplace transform r_p^∗(s). We define an observation interval as a period between when the previous call arrives and when the next call arrives (e.g., the interval [t2, t5] in Figure 6).

It is clear that the probability p described at the beginning of this section is the probability that no data session is in progress when the previous call is released (with probability p1) and no data session arrives before the next voice call arrives (with probability p₂).

The sequence of ts and tp forms an alternating renewal process [11], and therefore p1 =

E[tp]

E[tp]+E[ts]. Since a call release event is a random observer of ts and tp, p2 can be expressed as Pr[ta < τp]. According to the above description, we have

p = p₁p₂ =

 E[tp] E[tp] + E[ts]



Pr[ta< τp] (1)

Based on the inverse Laplace transform formula and the residue theorem [12, 13], Pr[ta<

τp] in (1) is derived as

Pr[ta < τp] = Z ^∞

τp=0

rp(τp) Z τp

ta=0

fa(ta)dtadτp

= Z ∞

τp=0

rp(τp)Fa(τp)dτp

=

 1 2πi

 Z σ+i∞

σ−i∞

Z ∞ τp=0

rp(τp)e^sτp f_a^∗(s) s

 dτpds

=

 1 2πi

 Z σ+i∞

σ−i∞

r_p^∗(−s) f_a^∗(s) s

 ds

= −X

z∈σ

Ress=zr_p^∗(−s) f_a^∗(s) s



(2)

where i = √

−1, σ is a sufficiently small positive number, σ^p is the set of poles of r_p^∗(−s) in the right half of the complex plane, and Ress=z denotes the residue at the pole s = z.

Alternatively, Pr[ta < τp] can also be derived as

Pr[ta < τp] = Z ∞

ta=0

fa(ta) Z ∞

τp=ta

rp(τp)dτpdta

= Z ^∞

ta=0

fa(ta)[1 − R^p(t)]dta

= 1 −

 1 2πi

 Z σ+i∞

σ−i∞

Z ∞ ta=0

fa(ta)e^sta

r_p^∗(s) s

 dtads

= 1 +X

z∈σa

Res_s=zf_a^∗(−s)

r_p^∗(s) s



(3)

where σa is the set of poles of f_a^∗(−s) in the right half of the complex plane.

For the demonstration purpose, we compute Pr[ta < τp] based on two cases of ta and tp

distributions.

Case 1. tais a Gamma random variable with the shape parameter k and the rate parameter λ, and tp is an Erlang random variable with the shape parameter m and the rate parameter γ. In this case, k > 0 is a real number and m ≥ 1 is an integer.

Case 2. tais an Erlang random variable with the shape parameter k and the rate parameter λ, and tp is a Gamma random variable with the shape parameter m and the rate parameter γ. In this case, k ≥ 1 is an integer and m > 0 is a real number.

The Gamma distribution is considered because this distribution is widely used in telecom modeling [14]-[17]. We also select the Erlang distribution because this distribution can be easily extended into a hyper-Erlang distribution, which has been proven to be a good approximation to many other distributions as well as measured data [13, 18]. The Laplace

transforms for the ta and the tp are

f_a^∗(s) =

 λ s + λ

k

and f_p^∗(s) =

 γ s + γ

m

(4)

For Gamma ta and tp, k and m > 0 are positive real numbers in (4). For Erlang ta and tp, k and m are positive integer numbers. From the residual life theorem [10] and (4), r_p^∗(s) is expressed as

r_p^∗(s) = γ sm

[1 − fp^∗(s)] = γ sm

 1 −

 γ s + γ

m

(5)

For case 1, we substitute (4) and (5) into (2) to yield

Pr[ta < τp] =

m−1

X

i=0 i

X

j=0

 γ^i−j mλ^i−j(i − j)!

  λ γ + λ

k+i−j i−j

Y

l=1

(k + l − 1) (6)

Note that in (6), when i − j = 0,

i−j

Y

l=1

represents an empty product, and its value is 1. From (1) and (6), p is re-written as

p =

m−1

X

i=0 i

X

j=0

 µγ^i−j

(γ + mµ)λ^i−j(i − j)!

  λ γ + λ

k+i−j i−j

Y

l=1

(k + l − 1) (7)

On the other hand, for case 2, from (4) and (5), (3) is re-written as

Pr[ta< τp] = 1 −

k−1

X

i=0

( γ λm−

i

X

j=0

 λ^j−1 j!mγ^j−1

  γ λ + γ

m+j j

Y

l=1

(m + l − 1) )

(8)

From (1) and (8), p is re-written as

p =

 mµ

γ + mµ

( 1 −

k−1

X

i=0

( γ λm−

i

X

j=0

 λ^j−1 j!mγ^j−1

  γ λ + γ

m+j j

Y

l=1

(m + l − 1) ))

(9)

Equations (7) and (9) are validated against the discrete event simulation experiments, which shows that the discrepancies between the analytic and simulation results are within 0.5%.

4 Numerical Examples

This section studies the call setup delays of DR and IR. Let tf be the time that the UE falls back from LTE to UMTS (i.e., Steps 1-10 in Figure 2). Let td be the UMTS outgoing call setup delay without the CS fallback (i.e., Steps 11-15 in Figure 2). Then the performance improvement α of the DR scheme over the IR scheme can be defined as

α = 1 − (1 − p)E[t^f] + E[td]

E[tf] + E[td] = pE[tf]

E[tf] + E[td] (10)

In (10), E[tf]+E[td] is the expected total call setup delay for IR, and (1−p)E[t^f]+E[td] is the expected total call setup delay for DR. The larger the α value, the better the performance of DR over IR. From the call setup delay measurement of Qualcomm (see Section 2), E[tf] = 2.5 seconds for 3GPP R8, E[tf] = 0.5 seconds for 3GPP R9, and E[td] = 4 seconds. From Huawei’s measurements [6], E[tf] = 9 seconds for 3GPP R8, E[tf] = 3 seconds for 3GPP R9, and E[td] = 5 seconds. We also measured the call setup delay at Broadband Mobile Lab of National Chiao Tung University [7], where E[tf] is more than 10 seconds and E[td] = 7 seconds. Our measurement results are more consistent with Huawei’s results than that of Qualcomm’s results. In this paper, we use Huawei’s results to compute α in (10).

We also note that although an LTE data connection is “always on”, the connection is in the idle mode (and is actually disconnected) if no data session is in progress. Because the expected session holding time is typically shorter than the expected inter-session arrival time [19, 20], we assume that 0.01E[tp] ≤ E[t^s] ≤ 0.1E[t^p]. We consider the effects of ts (the

session holding time), ta (the inter-call arrival time), and tp (the inter-session arrival time) on the probability p that a voice call can be connected without the CS fallback overhead.

Note that the voice call holding time tc does not affect p and is not considered. We also note that the α value is proportional to the p value (see (10)), and the effects on α are similar to those on p. Finally, to simplify our discussion, ts and ta are normalized by tp.

Effects of E[ts]/E[tp]: Figure 7 (a) shows that p decreases as E[ts]/E[tp] increases. When E[ts]/E[tp] increases, a call is more likely to be released in the ts interval. In this case, the UE will return to LTE immediately, and smaller p is observed. The non-trivial observation is that E[ts]/E[tp] has insignificant impact on p for all E[ts]/E[tp] values under our study. The probability p and the improvement α decrease by 8% when E[ts] increases from 0.01E[tp] to 0.1E[tp]. In other words, in this operational range, we can ignore the effect of E[ts]/E[tp] and can focus more on other parameters.

Effects of E[ta]/E[tp]: Figures 7 (a), 8 (a), and 9 (a) indicate that p decreases as E[ta]/E[tp] increases. When E[ta]/E[tp] increases, the data session is more likely to arrive before the voice call arrives (i.e., the UE will return to LTE before the next call arrives). Thus, a smaller p is observed. Figures 8 (a) and 9 (a) show that the effects of E[ta]/E[tp] become insignificant when Va or Vp is large, where large p and α are always observed.

Effects of Va: Figure 8 (a) indicates that p increases as Vaincreases. For a fixed E[ta] value, when Va increases, there are much more short ta intervals than long ta intervals. For short ta, it is very likely that ta < τp (i.e., larger Pr[ta < τp] is observed). From (1) and (10), p and α increase as Va increases.

Effects of Vp: Figure 9 (a) shows that p increases as Vp increases. When the inter-session arrival interval becomes more irregular (i.e., Vp increases), more long and short tp

intervals are observed. Since the call release events are more likely to fall in long tp

intervals and the next calls are likely to arrive before the next sessions arrive, larger p