initiation protocol (SIP) is studied. The transmissions on both the forward and reverse channel are assumed to experience Markovian errors. The session setup delay is evaluated for different trans-port protocols, and with the use of the radio link protocol (RLP). An adaptive retransmission timer is used to optimize SIP perfor-mances. Using numerical results, we find that SIP over user data-gram protocol (UDP) instead of transport control protocol (TCP) can make the session setup up to 30% shorter. Also, RLP dras-tically reduces the session setup delay down to 4 to 5 s, even in environments with high frame error rates (10%) and significant correlation in the fading process (fDT = 0.02). SIP is compared with its competitor H.323. SIP session setup delay with compressed messages outperforms H.323 session setup delay.
Index Terms—Correlated fading channels, IP-based wireless networks, radio link protocol (RLP), session initiation protocol (SIP), session setup delay, transport control protocol (TCP), user datagram protocol (UDP) .
I. INTRODUCTION
S
UPPORTING DATA services for applications like email, http, and Voice over internet protocol (VoIP) [1] in wire-less communication systems is gaining more importance with the development of 3G systems and beyond. The provisioning of such services requires that the system is robust enough to recover from errors that are likely to happen over radio chan-nels. Robustness can be provided by error control mechanisms such as automatic repeat request (ARQ) protocols, error correct-ing codes, and reliable transport protocols such as the transport control protocol (TCP).In this paper, we study the performances of the signaling protocol session initiation protocol (SIP) in terms of session setup delay. Developed by the Internet Engineering Task Force (IETF), SIP is “an application-layer control protocol that can es-tablish, modify and terminate multimedia sessions or calls” [2]. In wireless networks, two factors have a major impact on the performance of SIP session setup delay; namely, the physical channel, and the underlying protocols used by SIP. These two elements, usually studied separately, should be considered to-gether for an accurate evaluation of performance.
Manuscript received August 23, 2004; revised March 15, 2005, June 22, 2005. This work was supported in part by the Danish “Statens Teknisk-Videnskabelige Forskningsr˚ad” through the Center for Network and Service Convergence (CNTK) and in part by the Academy of Finland. The review of this paper was coordinated by Prof. D. O. Wu.
H. Fathi and R. Prasad are with the Center for TeleInFrastruktur (CTIF), Aalborg University, Niels Jernes Vej 12, 9220 Aalborg, Denmark (e-mail: hf,prasad@kom.aau.dk).
S. S. Chakraborty is with the Academy of Finland and Helsinki University of Technology, FIN 02015 HUT, Finland (e-mail: ssc@cc.hut.fi).
Digital Object Identifier 10.1109/TVT.2005.861213
the channel models used to evaluate the performance of proto-cols must reflect the physical layer characteristics in order to obtain meaningful results. Simplistic models may lead to unre-alistic results. In most of the models used to study session setup times, it was assumed that the errors were independent and iden-tically distributed (i.i.d.). Channel errors were assumed to be in-dependent from slot to slot, and the feedback channel was error-free, but the radio channels suffer from correlated multipath fad-ing with bursty frame errors. Errors are likely to be clustered, and to occur in bursts on both the forward and reverse chan-nels. This type of channel can be very well approximated by a two-state Markov model, also called a Gilbert–Elliot model [3]. In this paper, we particularly consider the performance of SIP over the reliable TCP, the unreliable user datagram pro-tocol (UDP), and an ARQ error recovery mechanism called radio link protocol (RLP). The purpose of RLP is to provide extra reliability to the Layer 2 on top of LLC, since LLC was primarily designed for wireline access networks, experiencing much less data losses. We investigate the impact of the FER and the burstiness of wireless links on the performance of SIP over these protocols. Note that to overcome losses, timers are set to trigger retransmissions of unacknowledged messages, and therefore permit detection of lost messages, but they increase the delay in case of lost messages. In this paper, we use an adaptive retransmission timer that is adjustable to the SIP transaction [4]. In most of the existing literature on session setup delay for VoIP, errors are assumed to be independent, and feedback to be error-free. In [5], the H.323 session setup time is evaluated over a wireless link with random errors and an error-free feedback channel. However, in [6], the performance evaluation using sim-ulations is done over correlated fading channels but it is limited to the TCP/RLP protocol stack. Also, a study exclusively on RLP performances using hidden Markov models for the fad-ing channels has been presented recently in [7]. Also, in [8], the delay of SIP messages induced by the transport protocols (SCTP, TCP, and UDP) is evaluated in a wired environment with very low FER. SIP session setup delay has been studied in [9] for the wired Internet. In contrast, here we focus on a wireless environment, and on the choice of protocol stack to optimize the VoIP session setup delay. Therefore, using the two-state Markov model, we evaluate the SIP/UDP/RLP stack and the SIP/TCP/RLP stack to compare their session setup delay under various degrees of burstiness.
The rest of the paper is organized as follows: we provide an overview of SIP, highlighting the packet loss recovery techniques, in Section II. Section III defines the session setup delay. In Section IV, the frame-level model for fading channels 0018-9545/$20.00 © 2006 IEEE
Fig. 1. SIP session setup.
is introduced. In Section V, a Markovian model is developed to study the performance of SIP over both TCP and UDP, with and without RLP, and in conjunction with the adaptive retransmission timer proposed in [4]. Results in Section VI compare SIP session setup delay over a variety of underlying protocols under different conditions of error rates and burstiness with H.323 performances.
II. OVERVIEW OFSIP PROTOCOL
The main purpose of SIP is to establish sessions between two user agents. A user agent contains both a client application that sends SIP requests, and a server application that accepts requests. SIP works together with the session description proto-col, which is in charge of describing the session to be opened. SIP messages can be carried by UDP or TCP [2].
A. SIP Session Establishment
SIP is a transactional protocol in the sense that a SIP trans-action consists of a single request and any responses to that request. The establishment of a session using SIP consists of an INVITE transaction and an ACK transaction. Fig. 1 illustrates the session setup between two user agents.
The calling party UAC A starts the transaction by sending an SIP INVITE request to the called party UAS B. The INVITE request contains the details of the type of session. When the called party decides to accept the call, a 200OK response is sent to show the agreement on the type of media. The final step is to confirm the media session with an ACK request. Then the media session is established.
B. SIP Reliability Mechanisms for Session Setup
To ensure the reliable delivery of the SIP requests and re-sponses involved in the INVITE and ACK transactions, retrans-mission mechanisms are needed at the user agent client and the user agent server.
1) Client Side: The client-side transaction consists of the
UAC sending the INVITE request and receiving the 200OK re-sponse. The UAC is aware of the successful transmission of the INVITE request as soon as it receives the 200OK response. If SIP messages are carried over UDP, the UAC retransmits the INVITE request after an interval that lasts Tr(1) s and dou-bles after each retransmission. The retransmissions cease upon the reception of a 200OK response at the UAC or after seven transmissions of the INVITE request.
Fig. 2. Session setup of SIP over TCP.
For TCP, there is no retransmission mechanism at the appli-cation layer; this is handled by the transport layer. TCP handles packet losses by setting a timer when it sends data, and if the data is not acknowledged when the timer expires, it retransmits the data [1]. But for any type of transport protocol, the retrans-missions of requests cease when the timer reaches 26· T r(1) s.
2) Server Side: The server-side transaction consists of the
UAS sending the 200OK response and receiving an ACK. The UAS is aware of the successful transmission of the 200OK and of the beginning of the call when it receives the ACK. The retransmission mechanism is identical to the one on the client side. In addition, each 200OK being received at the UAC triggers the retransmission of an ACK.
III. SESSIONSETUPDELAY
Several delays can be considered to assess the quality of ser-vice of signaling protocols. In this paper, we focus on the session setup delay defined as the period between the instant the UAC triggers the initiation with an INVITE request, and the instant the UAS has been alerted that the client received the server’s agree-ment upon the session (reception of ACK request at the UAS).
As the SIP session setup is the completion of the INVITE and ACK transactions, which can be seen as the client-side and server-side transactions, the SIP session setup delay is consid-ered as the cumulative delay of the completion of the transac-tions at the client side and at the server side.
It is important to note that provisional responses (i.e., 1xx) are not part of the message flow considered in this paper, because they do not influence the session setup delay. Only final re-sponses are important steps in the progress of the session setup. The underlying protocol used by SIP (UDP, TCP, and RLP) influences the session setup time. If SIP is used over UDP, the total session setup delay is the time needed for all messages involved in the setup (INVITE, 200OK, and ACK) to be suc-cessfully received by UAC (200OK) and UAS (INVITE, ACK). If SIP is used over TCP, as illustrated in Fig. 2, the total session setup delay is the addition of the setup time of the TCP session and the successful transmission time of all the SIP messages necessary to open a session.
Fig. 3. RLP scheme (1, 2, 3).
Fig. 4. Frame-level two-state Markov model.
SIP over RLP can reduce the effect of FER on the session setup time. RLP is a pure negative acknowledgement (NAK)-based selective repeat ARQ protocol. When the RLP receiver finds a frame in error, it sends back a NAK requesting the re-transmission of the erroneous frame. At the sender, each NAK received correctly triggers the retransmission of the frame re-quested in the NAK. RLP gives priority to RLP control frames first, then to retransmitted data frames, and finally to new frames. More information about RLP can be found in [10]. In this paper, we investigate RLP(1, 2, 3) that performs six retransmissions in three rounds of NAKs as illustrated in Fig. 4.
IV. FRAME-LEVELMODEL FORFADINGCHANNELS Rayleigh fading channels are usually mode using Markov models. In this paper, we focus on the two-state Markov model, also known as the Gilbert–Elliot model illustrated in Fig. 4.
In [11], it was shown that a fading channel can be well approx-imated by the Gilbert–Elliot model, whose transition probability matrix is given by M = 1− r r s 1− s
where (1− r) and s are the probabilities that the transmis-sion of frame i is successful given that the transmistransmis-sion of the frame (i− 1) was successful, or unsuccessful, respectively. These probabilities depend only on the average frame error rate p and the normalized Doppler frequency fDT , which
de-scribes the correlation in the fading process. fD is the Doppler
frequency and T is the frame duration. The average FER p corresponds to the steady-state probability that a frame error occurs:
p = r
r + s. (1)
Also, for a Rayleigh fading channel with a fading margin F , the average FER can be found as in [11]
p = 1− e−F1. (2)
The transition probabilities r and s are determined as in [11] by
s = 1− p
p · (Q(θ, ρθ) − Q(ρθ, θ)) (3) r = p
1− p· s (4)
where Q(x, y) is the Marcum-Q function, and
θ =−2 log(1 − p)/(1 − ρ2) (5)
ρ = J0(2πfDT ). (6)
J0is the Bessel function of the first kind and the zero order.
Note that the error bursts and error-free gaps are geometrically distributed, resulting in an average error burst length equal to 1/s. In [11] and [12], it was shown that the length of the frame er-ror bursts depends on the link fading margin and the normalized Doppler frequency. By varying the frame error rate p and the normalized Doppler frequency fDT , we can obtain fading
Fig. 5. Average length of the frame error bursts versus normalized doppler frequency.
Fig. 5 shows the average burst length as a function of the nor-malized doppler frequency and the fading margin F . When fDT
is small (<0.1), the process is very correlated, and we obtain long burst of frame errors. On the other hand, for larger values of fDT , the frame error bursts are very short, and successive
samples are almost independent.
V. PERFORMANCEANALYSIS OFSIP SIGNALING OVERWIRELESSLINKS
For our analysis, we consider a simple session setup messages flow of SIP as depicted in Fig. 1 for UDP and Fig. 2 for TCP. The following assumptions are made about the end-to-end SIP session:
1) The processing delays and queuing delays through the dif-ferent SIP servers (e.g., P/I/S-CSCF of the IP-multimedia subsystem in the UMTS) are not considered here. 2) TCP is assumed to operate in an interactive mode. The
delayed acknowledgement mode of TCP is turned off.
A. SIP Over UDP
1) Adaptive SIP Back Off Timer: The SIP back off timer after
the ith transmission T r(i) doubles after each retransmission. Hence
T r(i) = 2i−1· T r(1). (7) The initial retransmission timer T r(1) is a crucial parameter which should be optimized, since it has a direct impact on the session setup delay. It should not be too short otherwise, the packet is transmitted while a response is on the way to be re-ceived; and it should not be too long to avoid increasing the session setup unnecessarily if a loss occurs. Therefore, it has to be proportional to the transmission time of the messages in-volved in the transaction. It is function of the number of frames
k contained in the UDP datagram, of the end-to-end frame
propagation delay D, and of the interframe time τ , the time
interval between the transmissions of two consecutive frames. Let us consider the client-side transaction: transmission of the INVITE request (containing k1frames), and acknowledged by
the 200OK response (containing k2frames). Hence, the
retrans-mission timer of the client-side transaction is
T r(1) = D + (k1− 1) · τ + D + (k2− 1) · τ. (8)
For the server-side transaction, the value of T r(1) changes, reflecting the number of frames contained in 200OK (k2) and
ACK (k3).
2) Without RLP: We consider that the packet contains k
frames. Out of k frames, m frames are erroneous. m is the burst length, which depends on the degree of correlation and the FER. We assume that not more than one burst can affect a packet. We consider all the different positions of the burst in the packet and we assume that different cases corresponding to the different positions of the burst in the packet, are equiprobable. The packet error rate is the average rate over the packet error rates induced by the different positions of the burst in the packet. The packet error rate PER(k) is
PER(k) = k−m
j =0 1− ((1 − r)k−m −jsj)
k− m + 1 (9)
where j is the index reflecting the position of the burst in the packet.
The probability of retransmission q is the probability of a transaction having failed: this means that the first packet sent (INVITE request containing k1 frames) is lost, or that the
first packet is received but the response (200OK containing k2
frames) is lost. Therefore, the probability of having a retrans-mission of INVITE during a client-side transaction
q = PER(k1) + PER(k2)(1− PER(k1)) (10)
For the server-side transaction, the value of q changes, reflecting the number of frames contained in 200OK (k2) and ACK (k3).
Let N m be the maximum number of transmissions (for SIP,
N m = 7). The average (normalized) delay T tiUDPfor the suc-cessful transmission of the ith UDP datagram is
T tiUDP= 1 1− qN m · [(1 − q)(D + (k − 1)τ) + (1− q)q(T r(1) + D + (k − 1)τ) + (1− q)q2· (3T r(1) + D + (k − 1)τ) + · · · + (1− q)qN m−1· ((2N m−1− 1)T r(1) + D + (k− 1)τ)] = D + (k− 1)τ − T r(1) +(1− q)(1 − (2q) N m)) (1− qN m)(1− 2q) · T r(1). (11)
The total session setup delay is the addition of the delays for all the N messages necessary to set up a VoIP session using SIP over UDP. The average session setup delay T tUDPis
T tUDP= N
i=1
the missing frame. Finally in the third trial, three NAKs are sent triggering three consecutive retransmissions of the same missing frame.
For this analysis, the following terms need to be defined: 1) Xij = ith retransmission frame at the jth retransmission
trial received correctly at the destination.
2) Yij = ith NAK frame at the jth retransmission trial
re-ceived correctly at the source.
3) Cij = the first frame received correctly at the destination,
being the ith retransmission frame at the jth retransmis-sion trial.
4) Aj = the missing frame not received correctly at the jth
retransmission trial.
5) Bj = the missing frame not received correctly up to the
end of the jth retransmission trial.
If the ith retransmission frame at the jth retransmission trial is received correctly at the destination, the ith NAK frame at the jth retransmission trial has been received correctly at the source. The ith retransmitted frame is sent while the channel state is Good:
P (Xij) = 1− r. (13)
The ith NAK frame at the jth retransmission trial is received correctly at the source with probability s since all the previous frame were lost (Bad state of the channel):
P (Yij) = s. (14)
Therefore, if a frame is not received correctly at the jth retrans-mission trial, all the retransretrans-missions comprised in the jth trial are lost, then
P (Aj) = P ∩j i=1Xij∩ Yij = (1− (1 − r)s)j. (15) If the frame is aborted after the nth retransmission trial, it means that the frame is not received correctly up to the end of the nth retransmission trial; this is expressed as
P (Bn) = r(1− (1 − r)s) n (n +1)
2 . (16)
And, if the first frame received corresponds to the ith retrans-mitted frame of the jth trial, it means that the missing frame has been lost up to the (j− 1)th retransmission trial, and up to the (i− 1)th retransmissions in the jth trial.
P (Cij) = s(1− r)r(1 − (1 − r)s) j (j−1)
2 +i−1. (17) Therefore the probability of transmitting a frame successfully over the RLC layer is given by
P f = 1− P (Bn). (18) × n j =1 j i=1 P (Cij) D + (k− 1)τ + l 1− P (Bn) n j =1 j i=1 P (Cij) × (2j)D + j(j + 1) 2 + i τ . (19) We have maximum three RLP retransmission trials; i.e., n = 3. Considering that we are dealing with SIP transactions, the SIP timer also changes for the client side:
T r(1) = D1+ D2 (20) where D1 corresponding to the delay for INVITE (k1 frames)
and D2 the delay for 200OK (k2frames), and
q = 1− P fk2+k1. (21) For the server side, the timer is similar. The expressions of T ti
and T t remain unchanged.
B. SIP Over TCP
1) TCP Back Off Timer: TCP timer is based on
measure-ments of the round-trip time of a TCP segment. In this paper, we use the adaptive timer similar to the one used by SIP over UDP
T r(1) = 2D + (K1− 1) · τ + (K2− 1) · τ (22)
where K1is the number of frames contained in the data packet
(e.g., INVITE), and K2 is the number of frames contained in
the acknowledgement piggyback with the data to be sent by the receiver (e.g., 200OK).
2) Without RLP: For TCP, the SIP user agent follows TCP
specifications to retransmit messages until an acknowledgement is received. The TCP segments carry SIP messages in their payloads. The total session setup delay is the addition of:
1) The transmission delays of the messages necessary for establishing one TCP session (SYN/SYN-ACK/ACK). 2) The transmission delays for the three SIP messages
nec-essary to set up a VoIP session
a) the INVITE request is sent to the user agent server; b) the 200OK response is sent to the user agent client along with the TCP acknowledgement for the INVITE request;
c) the ACK request is sent to the user agent server along with the TCP acknowledgement for 200OK response; and
d) the TCP acknowledgement for ACK request is sent to the user agent client.
The probability of retransmission q is as mentioned in (10). The average delay for transmitting successfully an ith TCP seg-ment is analogous to the one for UDP expressed in (11). The total session setup delay is
T tTCP= N
i=1
T tiTCP (23) where N is the total number of TCP segments necessary to establish the session.
3) With RLP: For RLP(1, 2, 3), the expressions (19), (20),
and (21) are derived, respectively, for the increase of propa-gation delay, the retransmission timer, and the probability of retransmission.
VI. NUMERICALRESULTS
This section presents the results of the average session setup delay for SIP over different transport and radio link protocols. The model introduced in Section V implies that the average SIP session setup delay increases exponentially with the FER. The number and the size of the messages exchanged affect the aver-age session setup delay. The reduction of these factors leads to a shorter session setup delay. For the evaluation, the approximate size for each SIP message is obtained from packets captured by the protocol analyzer Ethereal [13] in our experimental testbed. Also, as stated in [14], SIP messages should be compressed us-ing mechanisms in [14] and [15] when conveyed via low-rate connections. Using basic compression with static dictionary, the SIP messages for the initiation are reduced by 36% with a de-compression memory size of 2048 B,1resulting in a processing
time of 30 ms per message on a 100 MHz processor [16]. The bit rate considered is 9.6 kbps.2
The session setup time depends not only on the average FER, but also on the amount of burstiness in the channel. The session setup delay is therefore computed for different degree of correla-tion: highly correlated errors (fDT = 0.02, see Fig. 5), weakly
correlated errors (fDT = 0.08), and almost independent errors
(fDT = 0.2). These results are compared to the session setup
delay obtained with an i.i.d. error process. The values of the delay D and the interframe time τ are set as in [5], respectively, as 100 ms and 20 ms. For SIP over TCP and UDP, the maximum number of transmissions N m is set to seven according to the specifications.
A. SIP Over UDP
Table I shows the size of the UDP datagrams and the number of frames per datagram. We assume that each UDP datagram is
1Which is the default size in the standard for low-end terminals.
2It is reasonable to assume that such bandwidth will be allocated for SIP
signaling in 3G systems even if 3G systems offer bandwidth as high as 10 Mbps. Of course, if higher bandwidths are used to setup the session, the delay is drastically reduced. However, as bandwidth is a scarce resource, it is more relevant to investigate smaller bit rates to highlight the factors involved in the optimization techniques. Also, the model used here is easily extendable for variable bit rates if the FER is fixed.
TABLE I
MESSAGESSIZE ANDNUMBER OFFRAMES FORSIP OVER
UDP SESSIONSETUP
Fig. 6. Average session setup delay in 9.6 kb/s channel for SIP over UDP with/without RLP(1, 2, 3).
Fig. 7. Average session setup delay in 9.6 kb/s channels for SIP over UDP with RLP and different degrees of correlation.
carried over one IP packet. The overall header is assumed to be 28 B (20 B for the IP header, and 8 B for the UDP header).
The total average session setup delay using the adaptive re-transmission timer is shown in Fig. 6. The average session setup delay is evaluated at various FER between 0–10%. However, the session is established for VoIP and voice services are supported if the FER is between 1% and 3%.
Fig. 6 shows that the session setup delay for different degrees of correlation are converging for high FERs (8–10%). For low
Fig. 8. Average session setup delay in 9.6 kb/s channels for SIP over TCP with/without RLP.
FERs (<3%), the longest delay is obviously obtained when the errors are highly correlated. The more the errors are correlated in the channel, the less FER is needed to reach the maximum session setup delay.
RLP improves the delay of the session setup considerably. In Fig. 6, it can be observed that RLP provides the minimum session setup delay. RLP limits the impact of the error corre-lation on the session setup delay: even for 10% FER and high correlation, the session setup delay is around 3.75 s, as shown in Fig. 7 (which is a zoom of the RLP part of Fig. 6). Also, Fig. 6 shows that the difference in the performances obtained for weakly correlated and for almost independent errors is rela-tively insignificant (100 ms) compared to the performance ob-tained for highly correlated fading. Moreover, there is almost no difference between results with low correlation (fDT = 0.2)
and random errors (fDT = 1).
B. SIP Over TCP
Table II shows the size of the TCP segments and the number of frames per segment. We assume that each TCP segment is carried over one IP packet. The overall header is assumed to be 40 B (20 B for IP header and 20 B for TCP header). Fig. 8 illus-trates the average session setup delay with different correlation
Fig. 9. Average session setup delay in 9.6 kb/s channels for SIP over TCP with RLP and different degrees of correlation.
Fig. 10. TCP versus UDP for SIP session setup delay in a 9.6 kb/s channel.
channels. The highest correlation gives the maximum session setup delay even for quite low FER (<0.5%).
The use of RLP results in a more optimized session setup delay. As shown in Fig. 9 (which is a zoom of the RLP part of Fig. 8), for high FER (10%) and high correlation (fDT = 0.02),
the session setup delay is around 4.75 s.
The session setup delay of SIP over TCP follows the same behavior as UDP, but the session setup delay for TCP is around 50% higher than that for UDP. Moreover, the slope of session setup delay for TCP is smoother than the one for UDP.
C. UDP Versus TCP
In case of i.i.d errors, Fig. 10 shows that if the FER is less than 4%, the session setup delay of SIP over TCP is as long as that over UDP. This is due to the use of the adaptive timer. For a
Fig. 11. H.225 call signaling message flow [5].
FER higher than 4%, TCP gives a delay 20% longer due to the TCP session setup. This phenomena disappears when the errors are more correlated. Indeed, for weakly and highly correlated errors, the session setup delay for TCP is about 50% higher than for UDP, almost independently of the FER. However, we notice that the more the errors are correlated, the more the session setup delay for TCP tends to overpass the one for UDP, even at low FERs.
D. SIP Versus H.323
H.323 [7] is the concurrent of SIP for establishing VoIP ses-sions. The session setup delay is very large in H.323 for a regular call which involves multiple messages from its underlying pro-tocols like H.225 [18] and H.245 [19]. Figs. 11 and 12 describe the message flows in H.225 and H.245 respectively. Both the H.225 and H.245, messages are transmitted over a reliable trans-port protocol TCP. Two different TCP sessions are established for H.225 and H.245 procedures.
In this section, we compare SIP and H.323 protocols using the adaptive retransmission timer and the FER model described previously. The size of the H.323 messages needed to compute the session setup delay are taken from [5]. The calculations are made for the regular H.323 session setup. The maximum number of transmissions allowed is 10 [5].
In case of i.i.d errors, SIP outperforms H.323 as shown in Fig. 13. Moreover the H.323 session setup time grows exponen-tially because ten retransmissions are allowed instead of seven retransmissions for SIP. This comparison shows the consider-able influence of the timer and of the maximum number of retransmissions allowed on the delay performances of the sig-naling protocol. For a fixed timer, the only influence on the delay is the number of messages. Fig. 14 shows the results obtained for a retransmission timer fixed to 1.5 s for all protocols. H.323 gives higher delay than SIP for any FER.
VII. DISCUSSION
Traditional error recovery mechanisms use retransmissions with a predefined fixed timer. When a fixed timer is used, the
Fig. 12. H.245 message flow [5].
Fig. 13. SIP versus H.323 over TCP in 9.6 kb/s channel with adaptive timer.
number of messages necessary to open a session is the main influencing factor on the session setup delay. But with the adap-tive timer, the size of the messages is as well, and is even a more significant factor than the number of messages. Therefore, to optimize further the SIP session setup delay, some compres-sion schemes have been used here such as the sign comprescompres-sion (SigComp) [14]. However, this method involves a tradeoff
be-Fig. 14. SIP versus H.323 over TCP in 9.6 kb/s channel with fixed timer (1.5 s).
tween the processing delay, and the transmission delay. The compression may shorten the transmission delay but may also lengthen the processing delay depending on the complexity of the compression scheme.
Another way of further improving the session setup delay over error-prone channels is through the implementation of error correcting codes such as the forward error correction (FEC). In FEC, redundancy is added at the transmitter and is used at the receiver to correctly recover the information, even in the presence of some transmission errors. But such technique marginally improves the performance due to the correlation of the channel. To overcome this issue, the channel has to be made memoryless through the use of an interleaver. However, this introduces additional delay to the system. Hybrid solutions (ARQ + FEC) may be a good combination to optimize the session setup delay.
VIII. CONCLUSION
In this paper, using a Markovian model to capture the bursti-ness of the channel, we have evaluated the average SIP session setup delay depending on the FER, and the burstiness of the wireless link. To use UDP instead of TCP can make the ses-sion setup 30% shorter. Low layer retransmisses-sion mechanisms such as RLP considerably improve the session setup delay. In environments with high FER, the session setup delay with RLP remains small (in the order of 4 to 5 s), even with a highly correlated error process.
With the adaptive timer, SIP gives a shorter delay than H.323 for FERs higher than 1% for i.i.d errors and for weakly corre-lated errors. The adaptive timer is efficient for optimizing the performance of protocols in general. But the performance of SIP using the adaptive timer has been improved by using SigComp to reduce the size of the SIP messages.
Also, error correction mechanisms or hybrid ARQ schemes could improve the performance of VoIP session setup time by
[2] J. Rosenberg, H. Schulzrinne, G. Camarillo, J. Peterson, R. Sparks, M. Handley, and E. Schooler, “SIP: Session Initiation Protocol,” Request for comments 3261 Intenet Engineering Task Force, Jun. 2002. [3] E. Gilbert, “Capacity of a burst-noise channel,” Bell Syst. Tech. J., vol. 39,
pp. 1253–1266, Sep. 1960.
[4] H. Fathi, S. Chakraborty, and R. Prasad, “Optimization of VoIP session setup delay over wireless links using sip,” in Proc. IEEE GLOBECOM, vol. 6, Dallas, TX, Dec. 2004, pp. 4092–4096.
[5] S. Das, E. Lee, K. Basu, and S. Sen, “Performance optimization of VoIP calls over wireless links using H.323 protocol,” IEEE Trans. Comput., vol. 52, no. 6, pp. 742–752, Jun. 2003.
[6] A. Chockalingam and G. Bao, “Performance of TCP/RLP prtocol stack on correlated fading DS-CDMA wireless links,” IEEE Trans. Veh. Technol., vol. 49, no. 1, pp. 28–33, Jan. 2000.
[7] Q. Gao, “Performance analysis of RLP over fading channels in CDMA systems,” in Proc. IEEE GLOBECOM, 2003, pp. 966–971.
[8] G. Camarillo, R. Kantola, and H. Schulzrinne, “Evaluation of transport protocols for the session initiation protocol,” IEEE Network, vol. 17, no. 5, pp. 40–46, Oct. 2003.
[9] T. Eyers and H. Schulzrinne, “Predicting internet telephony call setup delay,” in Proc. IP Telephony Workshop. Berlin, Germany, Apr. 2000 [10] “Data service options for spread spectrum systems: Radio link protocol
type 3,” 3GPP2, Tech. Rep. C.S0017-010-A, Jun., 2004. [Online]. Avail-able: www.3gppp.org
[11] M. Zorzi, R. Rao, and L. B. Milstein, “On the accuracy of a first-order Markov model for data block transmission on fading channels,” in Proc.
4th IEEE Int. Conf. Universal Personal Communication (ICUPC), Tokyo,
Japan, Nov. 1995, pp. 211–215.
[12] W. C. Jakes, Microwave Mobile Communications, N. Ed. New York: Wiley, 1974.
[13] Ethereal. [Online]. Available: http://www.ethereal.com
[14] R. Price et al., Signalling Compression, Request for Comments 3320, Internet Engineering Task, Jan. 2003.
[15] H. Hannu et al., “Signalling compression—Extended operations,” RFC 3321, Jan. 2003.
[16] M. Nordberg et al., “Improving sigcomp performance through extended operations,” in Proc. IEEE Veh. Tech. Conf., Orlando, FL, vol. 5, pp. 3425– 3428, Oct. 2003.
[17] Telecommunication Standardization Sector of ITU, “ITU-T Recom-mendation H.323—Packet based multimedia communications systems,” Orlando, FL, Feb. 1998.
[18] Telecommunication Standardization Sector of ITU, “ITU-T Recom-mendation H.225.0—Media stream packetization and synchronization on non-guaranteed quality of service LANs,” Nov. 1996.
[19] Telecommunication Standardization Sector of ITU, “ITU-T Recommen-dation H.245—Control protocol for multimedia communication,” Feb. 1998.
Hanane Fathi (S’03) received the M.S. degree in
electrical engineering from Aalborg University, Aal-borg, Denmark, and the telecommunications engi-neering Diploma from Ecole Centrale d’Electronique of Paris, Paris, France, both in 2002. She is now pur-suing the Ph.D. degree with the Center for TeleIn-frastuktur, Aalborg University.
Her research interests include VoIP in wireless heterogeneous systems, mobility management, sig-naling, Markov modeling of ARQ schemes, and au-thentication schemes in wireless systems.
Shyam S. Chakraborty received the M.Tech. degree
from the Indian Institute of Technology Delhi, New Delhi, India, and the Licenciate of Technology and D.Sc. (Technology) degree from Helsinki University of Technology (HUT), Espoo, Finland.
He is a Docent to the Department of ECE, HUT. He has been a Visiting Professor at the Asian Insti-tute of Technology, Guest Professor at Aalborg Uni-versity, Aalborg, Denmark, and Guest Researcher at TU-Berlin, Berlin, Germany. His research interests are Markov modeling of MAC and ARQ schemes, multihop ad hoc networks, diversity combining, VoIP in wireless systems, mo-bility management, etc. Since June 2005, he has been with Ericsson Finland.
Dr. Chakraborty is a Guest Editor of a special issue of the IETE Journal
of Research on “Protocols for resource, link and mobility management,” Guest
Editor of the IEEE JOURNAL OFSELECTEDAREAS INCOMMUNICATIONSspecial issue on “Mesh Networks” and General Co-Chair of the Workshop “Meshnets,” 2005. He received the Academy Fellowship from the Academy Finland in 2000.
Ramjee Prasad (SM’90) was born in Babhnaur
(Gaya), Bihar, India, on July 1, 1946. He received the B.Sc. (Eng.) degree from Bihar Institute of Tech-nology, Sindri, India, and the M.Sc. (Eng.) and Ph.D. degrees from Birla Institute of Technology (BIT), Ranchi, India, in 1968, 1970, and 1979, respectively. Since June 1999, he has been with Aalborg Univer-sity, Aalborg, Demark, where he is currently Director of Center for Teleinfrastruktur (CTIF), and holds the Chair of Wireless Information and Multimedia munications. He is coordinator of the European Com-mission Sixth Framework Integrated Project MAGNET (his personal Adaptive
Global NET) and was involved in the European ACTS project Future Radio Wideband Multiple Access Systems (FRAMES) as a DUT project leader. He is a project leader of several international, industrially funded projects. He has published over 500 technical papers, contributed to several books, and has au-thored, coauau-thored, and edited 11 books. In addition, he is the Coordinating Editor and Editor-in-Chief of the Kluwer International Journal on Wireless
Per-sonal Communications and a member of the editorial board of other international
journals, including the IEEE COMMUNICATIONSMAGAZINEand IEE Electronics
Communication Engineering Journal. He is also the founding chairman of the
European Center of Excellence in Telecommunications, known as HERMES. Dr. Prasad is a Fellow of the IEE, a Fellow of the IETE, a member of The Netherlands Electronics and Radio Society (NERG), and a member of IDA (Engineering Society in Denmark). He is advisor to several multinational com-panies and has served as a member of advisory and program committees of several IEEE international conferences.
