ACCEPTED FROM OPEN CALL
Voice over MPLS Compared to Voice over Other Packet Transport Technologies David Wright, University of Ottawa
ABSTRACT Most major carriers are deploying MPLS and are also migrating voice traffic to packet transport. This article analyzes how those trends can be combined. It compares the advantages of VoMPLS with VoATM, VoIP, and VoFR, from the viewpoints of bandwidth utilization, implementation issues, and the region of the network (access/backbone) in which implementation takes place. VoMPLSoPPP is more efficient than VoMPLSoATM or VoMPLSoEthernet. In the network backbone VoMPLSoPPP is most efficient. VoAAAL2oATM has intermediate efficiency, and VoIP is highly inefficient. In the access network use of header compression improves the efficiency of VoIP but does not bring it to the level of VoAAL2oATM, which is approximately as efficient as VoMPLSoEthernet. VoMPLSoPPP remains most efficient.
INTRODUCTION Two trends motivate this article: first, the growth of telecommunications industry interest in the implementation of multiprotocol label switching (MPLS) as an efficient transport technology, and second, the migration of voice to packet transport. Together they raise the issue of the relative advantages of MPLS over other packet technologies for voice transport. The MPLS Forum specification [1] details the encapsulation of voice directly over MPLS and competes with alternative specifications for voice over asynchronous transfer mode (VoATM) [2], voice over IP (VoIP), and voice over frame relay (VoFR) [3]. Much has been written about VoATM, IP, and FR, including the current author’s book [4]. This article evaluates how VoMPLS compares with voice transport over those other technologies. The comparative evaluation of VoMPLS from the viewpoint of bandwidth utilization is described first. Implementation issues and the region of the network in which implementation takes place are then presented. A summary is also given. Of primary concern in assessing transport efficiency is the number of voice samples or
124
0163-6804/02/$17.00 © 2002 IEEE
packets that can be multiplexed in a single transport packet. This includes the multiplexing of packets from a single voice connection and the multiplexing of several voice connections. We use the following notation: n = number of octets in a single voice data unit M = number of voice data units from a single connection that are multiplexed into a transport packet N = number of active voice connections generating packets in a single transport packet It is important to note that N is not the total number of voice connections. It is only the number of voice connections actively generating packets at the time a particular transport packet is assembled. If silence removal is operating, removing a fraction, f, of the voice data units, N active connections corresponds to a total number of voice connections = N/f. A typical value is f = 0.6, giving a total number of voice connections of 1.67N, of which, on average, 60 percent are silent at any given time.
VOICE OVER MPLS In this article VoMPLS refers to the transport of voice directly over MPLS [1] without any intervening layer such as IP. Figure 1 illustrates the data structure, in which voice is transported in primary subframes consisting of a VoMPLS header followed by voice. This header includes a channel identifier that allows the multiplexing of multiple voice connections within a single MPLS data unit, and a length indicator that allows multiple voice samples or voice packets from a single connection to be transported within a single MPLS data unit. The number of octets of each portion of the data structure is indicated in Fig. 1, using the above notation adapted to the MPLS situation: n = number of octets in a single voice data unit M = number of voice data units from a single connection multiplexed into a single primary subframe N = number of primary subframes per MPLS data unit
IEEE Communications Magazine • November 2002
Repeat N times
Voice/ MPLS/ PPP
1
4
Flag
PPP
4
4
Outer MPLS
4
Inner MPLS
Mn
VoMPLS
MPLS shim
Voice
2
1
FCS
Flag
Primary subframe Repeat N times
Voice/ MPLS/ Ethernet
8
14
Preamble
Ethernet
4 Outer MPLS
4
4
Mn
Inner MPLS
VoMPLS
2
Voice
CRC
Primary subframe
MPLS shim
Repeat N times 5 Voice/ MPLS/ ATM
ATM
4
4
4
Mn
Outer MPLS
Inner MPLS
VoMPLS
2
Voice
Pad
AAL5
Primary subframe
■ Figure 1. Data structures for voice over MPLS. Figure 1 includes three alternatives for the link layer: Point to Point Protocol (PPP), Ethernet, and ATM. In the case of PPP and Ethernet, a “shim” MPLS header is added to the data uni, [5], occupying 4 octets for each layer of the label stack. In this article we will focus on the situation of two layers as typical of current MPLS implementations, resulting in an outer and inner MPLS header in Fig. 1. For ATM, [6] specifies the use of ATM adaptation layer 5 (AAL5) encapsulation, and the use of either the virtual channel identifier (VCI) field or the VCI/virtual path identifier (VPI) field for the outer MPLS label. Even if these fields are used for the outer MPLS label, the outer MPLS header is also included within the MPLS data structure. Otherwise there would be no means of knowing whether to expect an MPLS header in that position. Thus, the use of the VPI/VCI fields for MPLS labels does not reduce the amount of MPLS overhead from that for PPP and Ethernet. The use of AAL5 implies the need to create a data unit that is a multiple of 48 octets, and padding is indicated in Fig. 1 for that purpose. MPLS and VoMPLS use 4 octet headers to facilitate 32 bit processing. Some voice codecs (e.g., G.729) produce voice packets that are not multiples of 4 octets in length. In these cases padding can be used to ensure that the next primary subframe will start on a 32 bit boundary. Alternatively, more than one voice packet can be multiplexed into the primary subframe so that it comes to a multiple of 4 octets in length. For instance, G.729 produces 10 octet packets, so 2 packets is a multiple of 4 octets. The analysis in this article uses this multiplexing approach, thus avoiding the necessity of padding.
VOICE OVER ATM Voice can be transported over ATM using AAL1 circuit emulation and AAL5, which are suited to constant bit rate voice and therefore do not bene-
IEEE Communications Magazine • November 2002
Repeat N times 5 Voice/ AAL2/ ATM
ATM
1 Start field
3
Mn
AAL2
Voice
Pad
CPS packet
■ Figure 2. The data structure for voice over ATM. fit from silence removal. Recent implementations use AAL2 [2, 7], which does not have this disadvantage; therefore, AAL2 is used in this article. Figure 2 illustrates the data structure for voice over AAL2 over ATM. The common part sublayer (CPS) packet consists of a 3 octet AAL2 header followed by voice code. The header contains a channel identifier for multiplexing several voice connections into a single ATM virtual circuit, and a length indicator that allows multiple voice samples or voice packets from a single connection to be transported within a single CPS packet. Several CPS packets are, in general, mapped into a single ATM cell and may overlap into the next ATM cell within the same virtual circuit. In order to inform the destination where the next (i.e., nonoverlapping) CPS packet starts within a given cell, a 1 octet start field is used in the 6th byte of the cell, with a pointer indicating the required position. Since ATM cells are fixed length, padding is required after the last CPS packet in order to fill out the payload. For VoATM our notation is: n = number of octets in a single voice data unit M = number of voice data units from a single connection multiplexed into a single CPS packet N = average number of CPS packets per ATM cell
125
Compress to 2-4
Voice/ RTP/UDP/ IPv4/PPP
1
4
20
Flag
PPP
IP
8 UDP
Repeat N times 12
2
RTP
RTP mux
Compress to 2-4
Voice/ RTP/UDP/ IPv6/PPP
1
4
40
Flag
PPP
IP
8 UDP
Mn
2
1
Voice
FCS
Flag
Mn
2
1
Voice
FCS
Flag
Repeat N times 12
2
RTP
RTP mux
■ Figure 3. Data structures for voice over IP.
Repeat N times 1 Voice/FR
2
Flag FR
1-3
1
Mn
Subframe
Voice Trans. Str.
Voice
2
1
FCS Flag
■ Figure 4. The data structure for voice over frame relay.
VOICE OVER IP In an IP network, voice is transported over Real Time Protocol (RTP), which is encapsulated in User Datagram Protocol (UDP) for which IP is the network layer. Either Point to Point Protocol (PPP) or synchronous optical network/synchronous digital hierarchy (SONET/SDH) can be used as the link layer. PPP adds a fixed amount of overhead per IP packet, whereas the SONET/SDH overhead per IP packet depends on the length of the IP packet. For simplicity, in this article we take VoIP to be a shorthand for VoRTPoUDPoIPoPPP. This structure, illustrated in Fig. 3, allows only packets from a single voice call to be transported within a single RTP packet, resulting in a significant amount of overhead. There are two approaches to reducing the proportion of overhead. Header compression can be used on a link-bylink basis to transport only those fields in the headers that change in an unexpected way from one packet to another. This can reduce the combined RTP/UDP/IP headers from 40–60 octets to 2–4 octets, as shown in Fig. 3. The application of header compression to RTP/UDP/IP is given in [8], which is suited to links with bandwidths up to T1/E1 and is used in this article. At higher bandwidths, compression results in an unrealistic amount of processing overhead at either end of the link. There have been several Internet drafts proposing methods for RTP multiplexing, which would allow packets from multiple calls to be transported within a single RTP packet, thus reducing overhead. However, these drafts have not yet progressed to become RFCs and have expired. An argument against RTP multiplexing is that in IP, multiplexing is already performed at the UDP layer, which contains port numbers for that purpose. However, only a single UDP packet can be transported within an IP packet, so UDP multi-
126
plexing does not result in a reduction of the considerable overhead shown in Fig. 3. It is of course possible to redefine IP so that a single IP payload can contain multiple UDP packets, but that would require a new IP protocol type. The vast installed base of current IP equipment makes it highly unlikely that there would be industry interest in such a modification to IP, and it is therefore not used in this article. The analysis presented below includes RTP multiplexing as an option, which is shown in Fig. 3 with 2 octets of overhead, since that was used in some of the expired Internet drafts. In summary, in this article VoIP uses RTP. It can optionally also use RTP multiplexing, for which there is no current standard. For VoIP, our notation becomes: n = number of octets in a single voice data unit M = number of voice data units from a single connection that are multiplexed into a single RTP packet N = number of RTP packets per UDP/IP/PPP packet
VOICE OVER FRAME RELAY VoFR [3] includes the multiplexing of several voice packets from a single call into a subframe, and several subframes from different voice connections within an FR payload (Fig. 4). The subframe header has one mandatory octet containing a subchannel identifier, to distinguish among voice connections, and two optional octets. There is one additional octet of voice transfer structure header, giving a sequence number and the codec type. For VoFR, our notation becomes: n = number of octets in a single voice data unit M = number of voice data units from a single connection that are multiplexed into a single subframe N = number of subframes per FR frame
BANDWIDTH UTILIZATION This section presents a comparison among VoMPLS and voice transport over other packet technologies from the viewpoint of efficient utilization of bandwidth. It evaluates the effect of the varying amounts of overhead indicated in Figs. 1–4 in terms of the percentage of band-
IEEE Communications Magazine • November 2002
kb/s
Voice data Unit length (ms) bytes (n)
Option 1 packing (M) [1], Annex A
Option 1 delay (ms)
Option 2 packing (M) delay = 30 ms
G.711 (PCM)
64
5
40
2
10
6
G.722 (SB-ADPCM)
64
5
40
4
20
6
G.723.1 (5.3kb/s)
5.3
30
20
1
30
1
G.723.1 (6.4 kb/s)
6.4
30
24
1
30
1
G.726 (ADPCM)
32
5
20
2
10
6
G.727 (EADPCM)
32
5
20
4
20
6
G.728 (LD-CELP)
16
2.5
5
4
10
12
G.729
8
10
10
2
20
3
■ Table 1. Voice codec characteristics.
VoMPLS/ ATM
VoMPLS/ PPP
VoMPLS/ Ethernet
VoAAL2/ ATM
VoIP4/ PPP
VoIP4oPPP with hdr comp
VoIP6 PPP
VoIP6oPPP with hdr comp
VoFR
Fixed overhead, F
21
16
34
6
48
11
68
11
6
Variable overhead, V
4
4
4
3
2
2
2
2
3
Vbl. overhead, V, 1 call
4
4
4
3
0
0
0
0
3
■ Table 2. Overhead for each transport option. width used for voice as opposed to overhead. This is done for the major voice codecs and varying numbers of voice connections per transport packet. Table 1 gives the characteristics of the voice codecs analyzed in this article, in terms of their bandwidth and the size of voice packet. G.711, G.722, G.726, and G.727, which are based on pulse code modulation (PCM) voice samples of a single byte, are shown with a voice packet of 5 ms, which represents a consensus from [1–3]. G.726 and G.727 can operate at a range of bandwidths, and we analyze the 32 kb/s option. Table 1 also indicates two options for the number, M, of voice data units from a single connection that are multiplexed into a single transport packet. The first option uses the default setting from Annex A of the VoMPLS specification [1], and the delay introduced by this option is also shown. The second option uses a value of M that produces equal delay, 30 ms, for each codec. Table 2 summarizes the overhead introduced by the data structures shown in Figs. 1–4. The variable overhead is the transport overhead per voice packet. The fixed overhead is the transport overhead independent of the number of voice packets. Since RTP multiplexing is not standardized it is treated as an option for the case of VoIP. In the case of a single voice connection, we assume RTP multiplexing would not be used, and a separate row in Table 2 specifies the overhead in this case.
A SINGLE VOICE PACKET PER TRANSPORT PACKET: N = 1 Tables 3 and 4 give the transport efficiency for a single voice connection, measured as
IEEE Communications Magazine • November 2002
(# voice bytes)/(# voice bytes + # overhead bytes) = NMn/(NMn + NV + F) where: V is the variable overhead. F is the fixed overhead. N, M, and n are defined in the previous section. Single voice connections are important in the access network where a customer uses separate connections for calls to different destinations. They result in single voice packets per transport packet. Below, we discuss multiple voice connections in a single transport packet, which is more suited to traffic between voice gateways in either the access network or the network backbone. Table 3 deals with option 1 in which the default number of voice packets, M, is given in Annex A of the VoMPLS specification [1]. Table 4 deals with option 2 in which the number of voice packets, M, gives a uniform packetization delay of 30 ms. It can be seen from Tables 3 and 4 that ATM is not an efficient transport option for a single voice connection, since a significant amount of padding must be added to the voice packets to fill up the ATM payload. Ethernet is also inefficient because of the high Ethernet overhead. This is acceptable in a LAN environment where bandwidth is low cost, but these figures indicate low efficiency for metro area Ethernet as an MPLS transport option. Tables 3 and 4 indicate that header compression adds significantly to VoIP transport efficiency; however, it should be borne in mind that header compression is suited only to links up to T1/E1 bandwidth, and is therefore mainly restricted to the access network. Frame relay is an equally efficient access network technology. VoMPLSoPPP is not as efficient as FR or VoIP in the access network. In general, VoMPLS is not bandwidth-efficient for
127
VoMPLS/ ATM
VoMPLS/ PPP
VoMPLS/ Ethernet
VoAAL2/ ATM
VoIP4/ PPP
VoIP4/PPP with Hdr Comp
VoIP6/ PPP
VoIP6/PPP with Hdr Comp
VoFR
G.711 (PCM)
75.5%
80.0%
67.8%
75.5%
62.5%
87.9%
54.1%
87.9%
89.9%
G.722 (SB-ADPCM)
75.5%
88.9%
80.8%
75.5%
76.9%
93.6%
70.2%
93.6%
94.7%
G.723.1 (5.3kb/s)
37.7%
50.0%
34.5%
37.7%
29.4%
64.5%
22.7%
64.5%
69.0%
G.723.1 (6.4 kb/s)
45.3%
54.5%
38.7%
45.3%
33.3%
68.6%
26.1%
68.6%
72.7%
G.726 (ADPCM)
37.7%
66.7%
51.3%
75.5%
45.5%
78.4%
37.0%
78.4%
81.6%
G.727 (EADPCM)
75.5%
80.0%
67.8%
75.5%
62.5%
87.9%
54.1%
87.9%
89.9%
G.728 (LD-CELP)
37.7%
50.0%
34.5%
37.7%
29.4%
64.5%
22.7%
64.5%
69.0%
G.729
37.7%
50.0%
34.5%
37.7%
29.4%
64.5%
22.7%
64.5%
69.0%
■ Table 3. Transport efficiency for option 1, that is, the number of voice packets given in Annex A of [1].
VoMPLS/ ATM
VoMPLS/ PPP
VoMPLS/ Ethernet
VoAAL2/ ATM
VoIP4/ PPP
VoIP4/PPP with Hdr Comp
VoIP6/ PPP
VoIP6/PPP with Hdr Comp
VoFR
G.711 (PCM)
90.6%
92.3%
86.3%
90.6%
83.3%
95.6%
77.9%
95.6%
96.4%
G.722 (SB-ADPCM)
90.6%
92.3%
86.3%
90.6%
83.3%
95.6%
77.9%
95.6%
96.4%
G.723.1 (5.3 kb/s)
37.7%
50.0%
34.5%
37.7%
29.4%
64.5%
22.7%
64.5%
69.0%
G.723.1 (6.4 kb/s)
45.3%
54.5%
38.7%
45.3%
33.3%
68.6%
26.1%
68.6%
72.7%
G.726 (ADPCM)
75.5%
85.7%
75.9%
75.5%
71.4%
91.6%
63.8%
91.6%
93.0%
G.727 (EADPCM)
75.5%
85.7%
75.9%
75.5%
71.4%
91.6%
63.8%
91.6%
93.0%
G.728 (LD-CELP)
56.6%
75.0%
61.2%
56.6%
55.6%
84.5%
46.9%
84.5%
87.0%
G.729
28.3%
60.0%
44.1%
56.6%
38.5%
73.2%
30.6%
73.2%
76.9%
■ Table 4. Transport efficiency for option 2, that is, the number of voice packets giving a 30 ms delay. transporting a single voice call, since for each VoMPLS option there is another transport technology that is more efficient. Comparing Tables 3 and 4, it is clear that efficiency can be improved at the expense of increased delay. Many packet voice networks aim for an end-to-end delay of 100 ms in order to achieve toll quality voice, and 30 ms packetization takes significantly from this delay budget. For the remainder of this article we therefore concentrate on option 1, the default VoMPLS packing.
MULTIPLE ACTIVE VOICE CONNECTIONS PER TRANSPORT PACKET: N = 1, …, 10 We now compare the efficiency of VoMPLS with other transport options as the number of active voice connections generating voice packets per transport packet is increased from N = 1 to N = 10. As remarked on earlier, if silence removal is removing a fraction, f, of voice packets, this corresponds to a total number of voice connections = N/f. In this section we deal with the number, N, of active (i.e., nonsilent) voice connections. The results are presented graphically in Figs. 5 and 6. Codecs are grouped according to the total number of voice octets from a single connection per transport packet. For instance, in Table 1, two 40-octet G.711 packets produce the same bandwidth efficiency as four 20-octet G.727 packets. In order to facilitate reading the graphs, only five
128
transport options are included per graph, with separate graphs for access and backbone options. Many transport options can be used in both parts of the network; in particular, VoAAL2oATM is included as both an access and a backbone option. Figure 5 gives the access options, and Fig. 6 gives the backbone options, with the vertical scale chosen for comparison over both options, that is, the vertical scale is the same on pairs of Figs. 5a and 6a, and so on. IPv6 is not included since it is similar to IPv4. In addition to the transport options given in Table 2, VoIP without RTP multiplexing is also shown in the graphs, since RTP multiplexing is nonstandard. We now present an analysis of Figs. 5 and 6. Above N = 10 active voice connections there is little difference (< 15 percent) in bandwidth utilization among the major transport options, and the other factors discussed in a later section dominate the choice of transport technology. N = 10 with 60 percent silence removal corresponds with about 17 voice calls between a given pair of voice gateways. In the access network, VoMPLSoEthernet does not compare well with VoFR and VoIP with header compression and RTP multiplexing, due to the high overhead associated with Ethernet. VoMPLSoPPP is intermediate between these two alternatives. VoMPLSoEthernet has similar bandwidth efficiency to VoAAL2oATM, which is a major digital subscriber line (DSL) access alternative.
IEEE Communications Magazine • November 2002
1
1 Bandwidth efficiency
Bandwidth efficiency
0.9 0.9
0.8
0.7
0.8 0.7 0.6 0.5 0.4 0.3 1
2 3 4 5 6 7 8 9 (b) Number of active voice connections
10
1
2 3 4 5 6 7 8 9 (d) Number of active voice connections
10
0.6 1
2 3 4 5 6 7 9 8 (a) Number of active voice connections
10 1 0.9 Bandwidth efficiency
Bandwidth efficiency
1
0.9
0.8
0.7
0.8 0.7 0.6 0.5 0.4
0.6
0.3 1
2
3 4 5 6 7 8 9 (c) Number of active voice connections
10
Bandwidth efficiency
1 0.9
VoMPLSoEthernet VoAAL2oATM VolP4oPPP hdr comp. RTP mux VoFR VolP4oPPP hdr. comp.
0.8 0.7 0.6 0.5 0.4 1
2 3 4 5 6 7 8 9 (e) Number of active voice connections
10
■ Figure 5. a) G.711 and G.727 in access; b) G.723 (5.3 kb/s) and G.729 in access; c) G.722 in access; d) G.723 (6.4 kb/s) in access; e) G.726 in access.
In the backbone network, VoMPLSoPPP is a clear leader in terms of bandwidth efficiency. Many carriers have a large installed base of ATM in the backbone and may wish to use it for voice transport, via either VoMPLSoATM or VoAAL2oATM. The latter is more efficient. Figures 5 and 6 show a sawtooth pattern whenever ATM is used for transport because of the need to add variable amounts of padding to fill out the fixed length ATM payload. This pattern is less pronounced in the case of G.722, since the amount of voice is large compared to the size of the ATM payload; and in the case of G.726, for which the amount of voice (40 octets) matches the ATM payload (48 octets), more closely than for the other codecs.
IEEE Communications Magazine • November 2002
Figures 5 and 6 illustrate clearly the inefficient use of bandwidth by VoIPv4, and the additional overhead in VoIPv6 makes it even less efficient. RTP multiplexing is clearly a must before VoIP can even start to compete with the other transport options. However, header compression is also necessary to make VoIP as efficient as VoFR, and this is only possible in the access network. In the network backbone, VoIP with RTP multiplexing but without header compression achieves performance around the average of the other options. In summary, MPLS is a very efficient transport option for voice, particularly if it runs over PPP. When MPLS is deployed over ATM it is less efficient than VoAAL2oATM.
129
1
Bandwidth efficiency
1
Bandwidth efficiency
0.9
0.8
0.7
0.8
0.6
0.4
0.2
0.6
11 11
99 33 55 77 (a) Number of active voice connections
99 33 55 77 (b) Number of active voice connections
Bandwidth efficiency
1
0.9 VoMPLSoATM VoMPLSoPPP VoAAL2oATM VolP4oPPP RTP mux VolPoPPP
0.8
0.7 11
99 33 55 77 (c) Number of active voice connections
1
1
0.9 Bandwidth efficiency
Bandwidth efficiency
0.9 0.8 0.7 0.6 0.5 0.4
0.8 0.7 0.6 0.5 0.4
0.3
0.3
0.2 11
33 55 77 99 (d) Number of active voice connections
11
33 55 77 99 (e) Number of active voice connections
■ Figure 6. a) G.711 and G.727 in backbone; b) G.723 (5.3 kb/s), G.728 and G.729 in backbone; c) G.722 in backbone; d) G.723 (6.4 kb/s) in backbone; e) G.726 in backbone.
IMPLEMENTATION ISSUES Most major carriers are deploying MPLS and are also migrating voice traffic to packet transport. This article analyzes how those trends can be combined. MPLS allows IP routing protocols such as Open Shortest Path First (OSPF), Border Gateway Protocol (BGP), and Intermediate System to Intermediate System (IS-IS) to be used on IP, ATM, FR, and optical equipment. It allows equipment to switch at the IP, ATM, FR, and optical levels, and therefore integrates IP, ATM, FR, and optical transport. Initial deployment of MPLS is in the network backbone where carriers have a mix of IP, ATM, and optical equipment. Later deployment is
130
in the access network where there is a mix of IP, ATM, and FR equipment. In the customer premises, MPLS is less necessary, since IP is the dominant technology and interworks well with Ethernet LANs. At the demarcation point between the customer premises and the public network, MPLS is implemented in integrated access devices (IADs) to transport the customers’ IP traffic over whatever technology the carrier has in place. Voice therefore enters the MPLS network at two points, the customer premises IAD and a public network voice gateway, as indicated in Fig. 7. The VoMPLS standard [1] does not specify signaling functionality for these gateways, whereas in IP,
IEEE Communications Magazine • November 2002
protocols such as MGCP and H.323 have been specified. The VoMPLS IADs and public network gateways packetize voice as described in an earlier section, giving rise to VoMPLS in the access and backbone networks, respectively. We first evaluate the advantages of VoMPLS in general and then give an evaluation specific to those two areas.
PSTN
In this section we compare VoMPLS with VoIP, VoATM, and VoFR from the viewpoints of features that apply anywhere in the network. Quality of Service — FR standards indicate how quality of service (QoS) is measured, but do not specify how a user or network operator can request QoS from the network. FR QoS is therefore implementation-dependent. In IP, QoS can be implemented using integrated services (IntServ) with Resource Reservation Protocol (RSVP) signaling [9], which allows the user to quantitatively specify an end-to-end delay (e.g., 100 ms). An alternative implementation uses differentiated services (DiffServ) [10], which allows the user to request a prioritization of the traffic, giving a qualitative specification of delay and loss behavior. IntServ requires network equipment to maintain knowledge of the state of each connection, making it unscalable to large networks. In ATM, the user can get a quantitative guarantee of delay and loss behavior on individual voice calls in a scalable way, making it a more comprehensive QoS implementation than IP. Also, ATM QoS is proven in the field in most public networks today. MPLS allows QoS to be associated with the label(s) so that all traffic within a label switched path (LSP) gets a specified QoS. In addition, QoS can be derived from the EXP field in the MPLS header, allowing the QoS to be finetuned for individual packets within an LSP. Thus, MPLS QoS provides comprehensive delay and loss behavior on aggregated traffic, making it ideally suited to voice transport. Signaling — FR switched virtual circuits (SVCs) have not been implemented, so signaling is not used. In IP, the signaling alternatives are H.323, MGCP, and MEGACO/H.248 for control of voice gateways from servers, and RSVP [9] for setting up a call across an IP network. The Session Initiation Protocol (SIP) can be used to locate the called party. As stated earlier, RSVP has scalability problems, so there is no suitable end-to-end signaling protocol for setting up voice connections across an IP network. Instead, the connectionless nature of IP is exploited allowing traffic to be prioritized using DiffServ, without the signaling overhead of end-to-end connection establishment. In ATM, the private network–network interface (PNNI) [11] allows connections to be set up with specified QoS, within which individual calls can be assigned different AAL2 channels. In MPLS, there are two alternative signaling protocols available: • LDP: Label Distribution Protocol, a streamlined protocol with fast LSP setup. An extension, CR-LDP [12], specific to MPLS is available. • RSVP: A protocol designed for IP [9], with scalability problems. An extension with traf-
IEEE Communications Magazine • November 2002
Voice gateway
PBX
IMPLEMENTATION IN GENERAL
VoMPLS
IAD
VoMPLS
Ethernet Customer premises
Public network
■ Figure 7. VoMPLS architecture. fic engineering, RSVP-TE, [13], specific to MPLS is available. In summary, MPLS and ATM provide more comprehensive options for QoS and signaling of aggregated voice traffic than FR and IP. MPLS and ATM provide similar capabilities. In addition, VoIP traffic can be transported over an MPLS network by interworking between DiffServ and MPLS.
IMPLEMENTATION IN THE ACCESS NETWORK Packet voice in the access network is currently transported over IP, ATM, and FR. IP has four advantages in the access network: • It is widely used in customer premises; therefore, voice may already be in IP packets when it gets to the IAD. • From the customer perspective, VoIP has the advantage that it can be integrated with other IP applications such as Web browsing and Web-based call centers. • In the case of a cable access network, IP is the standardized technology for transporting packet voice. • The access network generally uses links with low enough bandwidth that header compression can be used, adding significantly to bandwidth efficiency. ATM has the advantage that it is standardized for voice over DSL access using loop emulation, which allows voice to be transported over DSL to a Class 5 circuit switch without modifying the circuit switch. FR has the advantage that it is a widely used access technology for enterprise users, some of which use it to transport voice. By comparison, MPLS is a new technology for the access network. In principle, MPLS has the required technical capabilities to transport voice, and VoMPLS could be a software upgrade to a customer premises IAD. In practice, MPLS faces competition from VoFR on enterprise networks, from VoIP in the customer premises, and from VoATM loop emulation. There are two main drivers for VoMPLS in the access network: • MPLS virtual private networks (VPNs) as a service to customers • Widespread MPLS deployment in the network backbone, thus offering seamless VoMPLS end to end for customers
IMPLEMENTATION IN THE NETWORK BACKBONE Packet voice in the backbone network is currently transported over IP and ATM.
131
VoMPLS
VoIP
VoATM
VoFR
QoS
Comprehensive QoS on aggregated traffic.
Scalable with qualitative guarantees or non-scalable with quantitative guarantees.
Quantitative guarantees.
Not standardized.
Signaling
CR-LDP and RSVP-TE.
Non-scalable.
PNNI.
None.
Access network
Software upgrade to customer IAD.
Integration with CPE IP applications. Standard for cable networks
Standard for DSL loop emulation.
Widespread deployment.
Backbone network
Traffic engineering. Interworking with IP/DiffServ. Implementable over ATM.
No quantitative scalable QoS.
Large installed base. Proven network management
N/A
Bandwidth efficiency.
Very high with PPP.
Requires RTP multiplexing. Header compression improves efficiency in the access network.
Nonmonotonic with increasing traffic.
Very high
■ Table 5. Summary comparison of VoMPLS with packet voice alternatives.
IP has the advantage of simple operation, with no need to set up connections for voice calls and with a few DiffServ service classes giving preferential performance to voice traffic. ATM has the advantage of a large installed base of equipment in most public carrier networks, and well proven network management capabilities. MPLS offers: • The ability to interwork with IP, including DiffServ • Implementation over the existing installed base of ATM equipment • Sophisticated traffic engineering capabilities [14], enabling carriers to balance the load over their networks
CONCLUSIONS Initial deployment of MPLS is in the network backbone, which is an ideal location for implementation of VoMPLS. Voice constitutes the largest volume of public network traffic that needs stringent QoS, and MPLS offers comprehensive QoS. The large installed base of ATM equipment in the network backbone can be used to transport VoMPLS. When VoMPLS is implemented over PPP, the bandwidth efficiency is better than over ATM, making PPP an alternative when the functionality of ATM is not required. By comparison, VoATM has an intermediate level of bandwidth efficiency, and VoIP is highly inefficient unless RTP multiplexing is used. In the access network, VoFR is a widely accepted enterprise network solution. VoATM has the advantage that it is standardized for use over DSL. Bandwidth efficiency for VoIP in the access network is better than in the backbone, since header compression can be used. However, RTP multiplexing is also required in order for VoIP to be as efficient as the other transport technologies. VoMPLS is most efficient if PPP transport is used. VoMPLS faces more competition from VoIP, since it interworks seamlessly with customer premises VoIP deployment. In the same way that ATM deployment in the access network has followed its deployment in the backbone, we can expect a similar pattern for MPLS, offering seamless transport over backbone and access MPLS, interworking with
132
customer premises IP, and well managed, efficient QoS-enabled VPNs for customer traffic including voice. This article has analyzed the advantages of VoMPLS compared to voice transport over other technologies from the viewpoints of bandwidth efficiency, and implementation issues in the access and backbone networks. The major points are summarized in Table 5.
REFERENCES [1] MPLS Forum, “Voice over MPLS — Bearer Transport Implementation Agreement,” 2001 [2] ITU-T Rec. I.366.2, “AAL Type 2 Service Specific Convergence Sublayer for Trunking,” 1998 [3] Frame Relay Forum, “Voice over Frame Relay Implementation Agreement,” FRF.11, 1997. [4] D. J. Wright, Voice over Packet Networks, Wiley, 2001. [5] E. Rosen et al., RFC3032, MPLS Label Stack Encoding, 2001. [6] B. Davie et al., “MPLS Using LDP and ATM VC Switching,” RFC 3035, 2001. [7] ATM Forum, “ATM Trunking Using AAL2 for Narrowband Services,” 1999. [8] S. Casner and V. Jacobson, “Compressing IP/UDP/RTP Headers for Low-Speed Serial Links,” RFC 2508, 1999. [9] R. Braden, Ed., “Resource Reservation Protocol (RSVP) — Functional Specification,” RFC 2205, 1997. [10] S. Blake et al., “An Architecture for Differentiated Services,” RFC 2475, 1998. [11] ATM Forum, “PNNI: Private Network to Network Interface,” af-pnni-0055.000, v1.0, 1996. [12] B. Jamoussi et al., “Constraint-Based LSP Setup Using LDP,” RFC 3212, 2002. [13] D. Awduche et al., “RSVP-TE: Extensions to RSVP for LSP Tunnels,” RFC 3209, 2001. [14] D. Awduche et al., “Requirements for Traffic Engineering over MPLS,” RFC 2702, 1999.
ADDITIONAL READING [1] M. Handley et al., “SIP, Session Initiation Protocol,” RFC 2543, 1999.
BIOGRAPHY DAVID WRIGHT (
[email protected]) combines a Ph.D. in engineering from Cambridge University with his current position as full professor in the University of Ottawa, School of Management to provide a business perspective on telecommunications engineering. He is the author of the book Voice over Packet Networks, and his papers on voice and video over packet networks have been published by IEEE. He is included in Who’s Who in Science and Engineering, and provides training and consulting services to vendors and service providers in the telecommunications industry.
IEEE Communications Magazine • November 2002
