Вы находитесь на странице: 1из 7

A Comparison of In-Band and Out-of-Band Transport Options for Signaling

Malathi Veeraraghavan Haobo Wang


ECE Department ECE Department
University of Virginia Polytechnic University
Charlottesville, VA 22904 Brooklyn, NY 11201
mv5g@virginia.edu haobo_w@photon.poly.edu

Abstract: Signaling protocols for GMPLS networks have been within each SONET signal, and (ii) out-of-band signaling
standarized and are now being implemented. Most switch vendors channels, e.g., an Ethernet interface from the control (call) pro-
allow for signaling messages to be carried over in-band signaling cessor of the switch to an IP network.
channels as well as through out-of-band networks. In this paper,
we compare these two signaling transport options. In carrying out Service providers are given the flexibility to choose
this analysis, we allow for both software-implemented signaling between these two options. This work is intended to provide
protocol processors, as is common in most off-the-shelf switches some quantitative guidance to service providers on this ques-
today, as well as hardware-accelerated signaling protocol engines. tion of which option to use. In order to carry out a quantitative
Our research project on hardware-accelerated signaling is in sup- analysis, we need to characterize the signaling traffic load.
port of a service concept in which dynamically provisioned high- Since the signaling traffic load will be a critical parameter to
speed end-to-end circuits are used for single file transfers. In this determine the answer, we need to consider the applications
application, signaling message loads will be extremely high since expected for GMPLS signaling.
call holding times will be short. For such a usage, we need a sig-
naling transport solution that minimizes message transfer times. The two most commonly cited applications for GMPLS
Therefore this question of in-band versus out-of-band signaling networks are rapid provisioning and fast restoration. Rapid
transport is relevant to our research. Our analysis shows that provisioning is aimed at reducing the large delays incurred
with hardware-accelerated signaling engines, in-band signaling is today between when a customer requests a circuit, and when
the better option, while in a network with only software signaling the circuit becomes available for use. Many bandwidth-on-
processors, the out-of-band solution is better. demand service providers envision providing customers access
to servers with web interfaces that can accept requests for con-
Key words: Signaling, In-band, Out-of-band, Hardware-accel- nections and turn these into commands to the ingress node of
eration the connection (e.g., the customer’s router to/from which the
new connection is being requested). This ingress router would
I. INTRODUCTION then use GMPLS signaling to provision the connection through
The Generalized Multi-Protocol Label Switching a chain of SONET/SDH or WDM switches to the far-end
(GMPLS) architecture [1] and associated signaling protocols (egress) router specified in the customer’s request. The use of
for Synchronous Optical NETwork (SONET)/Synchronous GMPLS signaling and routing automates a process that cur-
Digital Hierarchy (SDH) and Wavelength Division Multiplex- rently needs manual intervention. While such requests will
ing (WDM) networks, such as Resource reSerVation Protocol- generate some signaling load, we expect this load to be rela-
Traffic Engineering (RSVP-TE) [2], have been defined to tively low because of the nature of the application. Enterprise
enable the deployment of dynamically controlled high-speed and/or ISP network administrators making such bandwidth-on-
circuit-switched networks. Equipment vendors for SONET/ demand requests to increase or decrease the capacity of their
SDH crossconnects and WDM optical crossconnects are enterprise wide-area access or inter-router links are not likely
implementing these protocols. Networks equipped with these to issue these requests too often. Furthermore call setup delays
signaling-enhanced circuit switches should be able to receive in the order of seconds should be sufficient for this application,
requests for OC-1 (~51Mbps ) and higher-rate circuits, and given that even with this order of setup delays, the overall ser-
respond with dynamically provisioned circuits. While much vice would be a vast improvement over the current service.
attention has been paid to the definition of signaling protocols, In the fast restoration application, GMPLS signaling is
little attention has been given to the transport mechanisms for used to establish replacement connections following a failure
sending signaling messages between circuit switches. The thus saving a service provider protection-path bandwidth. This
RSVP-TE specification [2] allows for the signaling messages application does require very low call-setup delays, if restora-
to be carried directly in IP packets or UDP datagrams, but it tion is to replace protection as a solution for handling failures.
does not constrain the links/paths used to transport these IP Furthermore, in this application, a large number of connections
packets in any way. may need to be re-established simultaneously, which implies
Most vendors have allowed for two options for the signal- that a large number of signaling messages could be generated
ing “links”: (i) in-band signaling channels, which are realized in a short time. Switches will thus need the ability to handle
as bandwidth set aside within user-plane interfaces between large call volumes.
two switches, e.g., the Data Communication Channel (DCC) We proposed a third application class, file transfers, for

1
GMPLS-enabled high-speed circuit-switched networks [3].
Our work on this usage of GMPLS networks, called Circuit- DCC

-12
OC
{
switched High-speed End-to-End Transport ArcHitecture OC-1
(CHEETAH) is currently being sponsored in an eScience
project by the NSF. We identified file transfers as being ideally
suited for high-speed circuits given that a file transfer can take
advantage of any data rate it is provided (the higher the better), (a) In-band signaling case 1 (b) In-band signaling case 2 (c) In-band signaling case 3
and furthermore, a file transfer requires the movement of
Fig. 2 Three cases of in-band signaling
stored data, which means there is no inherent burstiness in the
generation of data at the sender. The latter makes it ideal to face is partitioned into signaling bandwidth and user-plane
keep a circuit pipe fully utilized. However, file transfer times bandwidth. For example, one or more OC-1s can be set aside
can be quite short. For example, it takes only 800ms to trans- within each user-plane interface to exclusively carry signaling
fer a 100MB file on an 1Gbps circuit. Therefore this applica- messages (e.g. Fig. 2b shows an OC-1 signaling channel within
tion can be regarded as rather extreme in its generation of calls an OC-12 interface). In the second case, one or more OC-1s of
and corresponding signaling message load. The in-band/out-of- one user-plane interface between two switches, which are con-
band question becomes important in this application because of nected by N interfaces, are set aside for signaling, as shown in
its need to support high signaling loads and low setup delays. Fig. 2c. The disadvantage of these two alternatives is that there
is revenue loss due to allocating user-plane bandwidth to sig-
Call setup delay includes (i) signaling message processing naling.
delays, (ii) round-trip propagation delay, and (iii) signaling
message emission delays. Toward reducing signaling message Fig. 1b illustrates the out-of-band signaling option. The key
processing delays, in an NSF-sponsored project [4], we imple- difference is that the path taken by the signaling channel is nec-
mented a subset of RSVP-TE in hardware to reduce the signal- essarily different from the path taken by the user-plane inter-
ing message processing overhead to the order of microseconds. faces that it supports. For example, the signaling channel could
The second component, round-trip propagation delay, cannot pass through packet switches while the user-plane interfaces
be reduced due to speed-of-light constraints. The goal of this are direct (logical or physical) between two switches. In other
work is to reduce the third component, emission delays, with- words, the signaling channel is separate from and independent
out of course compromising utilization. This led us to compare of the data channel. A classical example of out-of-band signal-
in-band and out-of-band signaling transport architectures. ing is the Signaling System 7 (SS7) network, which is used to
carry signaling messages between DS0-based telephone circuit
Section II describes the in-band and out-of-band transport switches. SS7 is a connectionless packet-switched technology.
options. We describe our queueing models for the two options
in Section III, and provide numerical results of our comparison Unlike in telephone networks, when it was economically
in Section IV. Our conclusions are presented in Section V. feasible to create a dedicated connectionless packet-switched
network just for signaling traffic, in today’s environment, with
II. IN-BAND AND OUT-OF-BAND SIGNALING OPTIONS the ubiquity of the Internet, it is more likely that GMPLS-
enabled SONET/SDH/WDM circuit switches will leverage the
Fig. 1 illustrates the two options for signaling transport: in- Internet for signaling message transport. A service provider
band signaling and out-of-band signaling. could simply connect the control processors of their GMPLS
Signaling R2 R3
enabled circuit switches to the Internet and expect it to route
User
R1 IP network R6 signaling messages as needed. A potential drawback is latency.
SW2 SW3 R4 R5 We will create models for such out-of-band signaling channels
SW1 SW6
SW2 SW3 routed through the Internet, and analyze these models to pre-
SW1 SW6 dict performance.
SW4 SW5 SW4 SW5

(a) In-Band Signaling (b) Out-of-Band Signaling III. DELAY MODELS FOR THE TWO OPTIONS
Fig. 1 In-band and out-of-band signaling options
In this section, we set up and analyze models for the two
signaling transport options, in-band signaling and out-of-band
In the in-band signaling option, the signaling traffic shares
signaling. Our goal is to compute the total delay incurred in
the same channel as the data traffic. For example, the DCC
processing and sending a signaling message from one switch to
channel in SONET signal can be used to transport signaling
the next successfully. This consists of the following compo-
messages, as shown in Fig. 2a. However, the bandwidth of the
nents: (i) queueing delay plus service time at the signaling pro-
DCC channel is limited. For each separate OC-1 signal, the
cessor, (ii) queueing delay plus transmission time on the
Section DCC has a bandwidth of 192Kbps , and the Line DCC
signaling channel, (iii) total delay to successfully send the mes-
of 576Kbps .
sage on the signaling channel, which includes retransmissions
There are two alternatives that can overcome the bandwidth in case of errors. We list our assumptions and notation, and
limitation of the DCC channel. In the first case, a switch inter- then describe our queueing model.

2
A. Assumptions Table 1: Notation

We assume that the call arrival process is Poisson, which Symbol Meaning
has been shown to hold true for FTP applications in [6]. Each
call involves multiple signaling messages. For example, in qj Probability of sending an outgoing signaling message to the
th
RSVP-TE, each call may involve Path messages, Resv mes- j in-band signaling channel; j = 1, …, n
sages, and PathTear/ResvTear messages. We assume the arrival
T proc A random variable denoting the response time at the signaling
of signaling messages is also Poisson. Although the messages protocol processor (waiting time plus service time)
involved in a call are related, we argue that since the duration
T tx A random variable denoting the response time at the signaling
of a call is random, the arrival times of PathTear/ResvTear channel transmitter (waiting time plus service time)
messages are independent of those of Path and Resv messages.
Tn One-way network delay in sending a signaling message from
Besides, after a switch sends out a Path message, the Path
one switch to the next
message may go through a random number of downstream
switches before reaching the destination. The time to the recep- To Initial time-out value of the retransmission timer at the sender
tion of a Resv message following the transmission of a Path
p Probability of packet loss
message is thus randomized.
The service times at signaling processors and the transmis- C. Queueing model
sion times at signaling channels are both assumed to be deter-
ministic. The processing of different RSVP-TE signaling There are two servers related to signaling at a GMPLS-
messages takes different times, and Path message is the most enabled circuit switch: (i) signaling protocol processor, and (ii)
time consuming one. In [4], we introduced a pipelined archi- signaling channel transmitter. We assume that the switch has a
tecture for hardware-acceleration of RSVP-TE signaling mes- single signaling protocol processor irrespective of whether the
sage processing. In order to achieve full pipelining and the signaling link solution is in-band or out-of-band. We first
highest throughput, we purposely inserted dummy cycles when describe our queueing model for the signaling protocol proces-
processing Resv, PathTear, and ResvTear messages so that the sor and the signaling channel transmitter without consider-
processing of all four messages will take the same number of ations of message loss and subsequent retransmissions. We
clock cycles. Even without such an implementation, the differ- then improve this model by adding in the possibility of mes-
ence in processing times is not significant for different mes- sage loss and retransmissions.
sage type. As for the length of signaling messages, the Path 1) Model without retransmissions
and Resv messages are roughly equivalent in size, while the
PathTear and ResvTear are roughly equivalent in size. The q1 µ txIB 1

approximate size is a few 100 bytes. Therefore we assume that µ proc q 2 µ txIB 2

λ µ proc µ txOOB
signaling processing times and signaling transmission times qn µ IBn
tx
λ
are deterministic. n
(a) In-Band Signaling, where qi = 1 (b) Out-of-Band Signaling
We assume that all messages fit in one packet, since the i =1

maximum transmission unit size of most existing networks is Fig. 3 Queueing models of the signaling protocol processor and the sig-
naling channel transmitters at a GMPLS enabled circuit switch
larger than the size of an RSVP-TE message.
Fig. 3 illustrates our queueing models of these two servers
B. Notation
for both in-band and out-of-band solutions. In the in-band solu-
Our notation is shown in Table 1. In the rest of the paper, tion, there are n signaling channels given our assumption of n
the superscripts, OOB and IB, may be used on some of these GMPLS-enabled neighbors to the switch (see Table 1). In the
parameters as appropriate for out-of-band and in-band, respec- out-of-band solution, we assume that there is only one out-of-
tively. band signaling channel.
Table 1: Notation Given that the models shown in Fig. 3 are networks of
queues, we consulted the literature on modeling such queueing
Symbol Meaning networks. Most of the solutions for queueing networks, such as
Burke’s theorem and Jackson’s theorem, are for a network of
λ Aggregate signaling message arrival rate
M/M/1 queues while in our case, as stated in subsection A, the
µ proc Service rate of the signaling processor service times of both queues are deterministic.
Even though there is no analytical derivation for a network
µ tx Service rate of the signaling channel transmitter
of queues with deterministic service times (to our knowledge),
Number of GMPLS-enabled neighbors to a switch (also the it turns out that if we take into account practical considerations,
n
number of in-band signaling channels) this network of queues reduces to a single M/D/1 queue. Our
reasoning is as follows. The service rate of the signaling pro-

3
cessor, µ proc , is determined by the switch vendor, while the channel. In the in-band solution, a similar reasoning holds if
service rate of the signaling channel, µ tx , either in-band or the individual signaling channel rates are selected according to
out-of-band, is determined by the service provider. We expect the probabilities q j (see Fig. 3 and Table 1), i.e.,
the switch vendor to select µ proc so that the user-plane inter-
IBj
faces and the signaling protocol processor are equally utilized. µ tx = q j µ proc , j = 1, 2, …, n . Using the argument stated
Once this decision is made, and a switch is manufactured, before, we assume that the service provider will either choose
µ proc is fixed. A service provider who purchases this switch all the in-band signaling channel rates such that
then has only one degree of freedom to choose an appropriate IBj
bandwidth level for the signaling channels, i.e., the service pro- µ tx = q j µ proc , or choose rates that are much lower,
vider can select µ tx . IBj
µ tx « q j µ proc . Under these conditions, our earlier reasoning
The service provider should limit µ tx ≤ µ proc , because
choosing a µ tx larger than µ proc is an unnecessary waste of of treating the queueing network as a single queue holds.
bandwidth. Therefore we attempt to solve the network of 2) Model including retransmissions
queues model in Fig. 3 for only two cases, µ tx = µ proc , and In this model, we include the third component of the delay,
µ tx < µ proc . We further limit the second case to only which is T n , the one-way network delay (see Table 1). This
µ tx « µ proc . Reasoning that a service provider may choose delay is impacted by losses and retransmissions. Message loss
µ tx < µ proc for costs reasons, we further argue that in this is possible irrespective of whether the signaling channel is in-
case it is more likely that µ tx « µ proc because cost differen- band or out-of-band. The reasons for message loss in the in-
tials for minor drops in transmission rates are often not signifi- band case are due to bit- and burst-errors on the links, and
cant (e.g., compare the cost of an 800Mbps circuit vs. a 1Gbps receive-buffer overflows due to flow control problem. Bit- and
circuit). Clearly the service provider should recognize that in burst-link errors arise due to noise and interference on the
choosing this option instead of the µ tx = µ proc option, the physical media. Even though optical fiber, the physical
network is being planned for a lower call arrival rate than the medium of these high-speed circuit-switched networks, is
maximum rate for which the switch is designed. fairly reliable, link errors are unavoidable. Receive-buffer
For these two cases, µ tx = µ proc and µ tx « µ proc , the overflows will occur if the signaling protocol processor at the
queueing network model of Fig. 3 is reduced to a single M/D/1 sending switch is faster than that at the receiving side. Since
queue. This is because under our assumption (see subsection different switch vendors could use different implementation
A) that the service time at the signaling protocol processor is techniques for the signaling protocol processor, this flow con-
deterministic, the departure rate is no more than µ proc . If trol problem may arise. In the out-of-band solution, an addi-
µ tx = µ proc , the second server (signaling channel transmit- tional cause of message loss is buffer overflow at any of the
ter) will keep up with the arriving signaling messages. In other routers on the path between the sending and receiving
words, the second queue is always empty, and the delay at the switches.
second queue is solely the service time (or emission time). On Since RSVP-TE is specified to use UDP or raw IP, and
the other hand, if µ tx « µ proc , since λ ≤ µ tx for the system to these protocols do not offer a reliable service, RFC 2961 [5]
be stable, λ « µ proc , which implies that the first queueing sys- proposed an exponential back-off retransmission algorithm to
tem is lightly loaded. We therefore assume that when a signal- provide reliability. Each message has a unique MESSAGE_ID,
ing message arrives, it is served immediately and the first and a timer starts counting after the message has been transmit-
queue is always empty. ted. If the corresponding acknowledgment (MESSAGE_ID_
For these two cases, µ tx = µ proc and µ tx « µ proc , the ACK) is not received within a time-out threshold, the signaling
queueing network model of Fig. 3 is reduced to a single queue. message is re-transmitted, the corresponding time-out thresh-
The mean response time for a signaling message at a switch, old is doubled (starting from T 0 ).
E [ T sw ] , is given by:
λ Transmitter
(1-p)
1
E [ T sw ] = E [ T proc ] + ------- , if µ tx = µ proc (1)
µ tx p

f(T0)
1
E [ T sw ] = E [ T tx ] + ------------- , if µ tx « µ proc (2)
µ proc Fig. 4 Signaling channel transmitter model including retransmissions

where E [ T proc ] is the mean response time at the signaling Assuming that retransmissions will, in typical implementa-
protocol processor queue, and E [ T tx ] is the mean response tions, join the signaling channel transmitter queue, and that it is
time at the signaling channel transmitter queue (see Fig. 3). valid to rewrite (1) to be the same as (2) (in other words for
both cases, µ tx = µ proc , and µ tx « µ proc , we use (2)), we
The above reasoning applies directly to the out-of-band redraw the model for the second queue (the signaling channel
case (Fig. 3b) since in this case there is only one signaling transmitter) as shown in Fig. 4. With probability p , a message

4
is lost, which means a time-out occurs, and the message is where ρ = λ ⁄ µ tx . The first term, 1 ⁄ µ proc , is common to
retransmitted. To model exponential doubling, we show the
both in-band signaling and out-of-band signaling. Therefore,
time-out value as a function f of the initial time-out value T o .
p , µ tx , and T n determine which is better, in-band signaling or
The mean response time at the transmitter queue, E [ T tx ] is
a sum of the mean waiting time, E [ W ], and the service delay out-of-band signaling under different sets of conditions. We
(which is a constant, 1 ⁄ µ tx , given our assumptions). The consider different numerical values for these input parameters
mean waiting time, E [ W ], can be computed assuming that the and compare the two signaling transport solutions.
message arrival rate at the transmitter queue is given by:
IV. NUMERICAL RESULTS
2 λ 3
λ + λp + λp + λp + … = ---------------- (3)
(1 – p) A. Input parameter values

Table 2: Input parameter values


but for small p , this can be approximated to λ .
Since in RSVP-TE, the signaling protocol processor is also Symbol Value
involved in handling retransmissions, the same load appears at
both the signaling protocol processor and the signaling channel λ Varied from 0.01µ tx to 0.95µ tx
transmitter. Thus, rewriting (1) to be same as (2) is not affected
because of retransmissions. µ proc 200, 000 /sec, hardware signaling
200 /sec, software signaling
Using µ proc , T n , T tx , T 0 , and p (see Table 1), and the
retransmission algorithm described above, we determine the µ tx µ tx = 200, 000 /sec, hardware signaling, µ tx = µ proc
mean total time to process and send the outgoing signaling
µ tx = 40, 000 /sec, hardware signaling, µ tx « µ proc
message successfully to the next-hop switch as:
µ tx = 200 /sec, software signaling
E [ T ] = 1 ⁄ µ proc + T n + [ ( E [ T tx ] ) ( ( 1 – p ) + (4)
2 n 5 and 10 . For the purpose of comparison, the aggregate
2p ( 1 – p ) + 3p ( 1 – p ) + … ) ] + T 0 ( p ( 1 – p ) + bandwidths of in-band and out-of-band signaling are the same,
2 3
3p ( 1 – p ) + 7p ( 1 – p ) + … ) therefore µ tx
IBj OOB
= µ tx ⁄n

The explanation for the above equation is as follows. The qj Probability of sending an outgoing signaling message to the
first two terms, 1 ⁄ µ proc and T n , correspond to the processing j
th
in-band signaling channel; j = 1, …, n
delay and the time to transmit a message successfully one-way
from the sending switch to the receiving switch. The next term Tn
IBj 0.2ms in metro area
corresponds to the mean response time (mean waiting time 25ms in wide area
plus mean service time) waiting in the signaling channel trans-
Tn
OOB 0.5ms and 1ms in metro area
mitter queue. This delay is incurred once if the first transmis-
sion is successful (probability ( 1 – p ) ), twice if the first 35ms and 45ms in wide area
transmission is unsuccessful but the second transmission is a IBj IBj
success and so on. The last term in (4) corresponds to the time- To 3T n
out value, which is being doubled for every retransmission. OOB OOB
To 3T n
1 1 p
E [ T ] = ------------- + T n + ------------ E [ T tx ] + --------------- T 0 (5)
µ proc 1–p 1 – 2p p
IBj
10
–8
and 10
–4

OOB 1% and 5%
Assuming T 0 = 3T n (counting round-trip network delay p
and other minor delays), (5) can be re-written as:
1 1 1+p With hardware-accelerated signaling engine [4], RSVP-TE
E [ T ] = ------------- + ------------ E [ T tx ] + --------------- T n (6) signaling messages can be processed within several microsec-
µ proc 1 – p 1 – 2p
onds. Accordingly we assume µ proc is 200, 000 /sec for a
Using the mean response time of an M/D/1 queue for hardware-accelerated signaling engine and 200 /sec for a soft-
E [ Ttx ] , we have: ware signaling processor.
For hardware-accelerated signaling, when µ tx = µ proc ,
ρ
1 – --- µ tx is 200, 000 /sec; when µ tx « µ proc , we assume
1 1 2 1 1+p 1
E [ T ] = ------------- + ------------ ⋅ ------------ ⋅ ------- + --------------- T n (7) µ tx = --- µ , or 40, 000 /sec. For software signaling, even
µ proc 1 – p 1 – ρ µ tx 1 – 2p 5
when µ tx proc= µ proc , µ tx is 200 /sec or 200kbps (assuming
125 bytes or 1000 bits per message). Considering the fact that

5
even a single SDCC channel has a bandwidth of 192kbps , IBj
delay is negligible. For in-band signaling, the values of p
assigning µ tx a value less than 200 /sec is impractical. There-
–8 –4
fore we do not consider µ tx « µ proc for software signaling. we use are 10 and 10 (see Table 2). In Fig. 5 and all fol-
In order to choose reasonable values for T n and p , we refer lowing figures, we cannot observe notable difference between
to the National Laboratory for Applied Network Research in-band signaling cases of p
IBj
= 10
–4
and p
IBj
= 10
–8
.
(NLANR)1 for Round-Trip Time (RTT) and packet loss rate
measurements. We randomly chose North Carolina University Fig. 6 shows the total mean delay in the wide area. Here,
(NCSU) as our starting point. For the data collected on June the total delay is even more dominated by T n , which is signifi-
14, 2004, we noticed that the RTT between NCSU and the cantly higher than the first two terms of (7), and hence in-band
North Carolina GigaPOP, is 1ms . We used the corresponding signaling outperforms out-of-band signaling.
one-way delay, 0.5ms , for our metro setting. The distance
between these two locations is 20.6 miles. The propagation
delay is around 0.16ms . We also notice that in a similar geo-
graphical distance (22.6 miles to be accurate), the RTT
between NCSU and University of North Carolina, Chapel Hill
(UNC) is 2ms , or 1ms for one-way. The different RTTs may
result from different network conditions. Based on the data
OOB
obtained from NLANR, we choose 0.5ms and 1ms for T n
IBj
in the metro area. We choose 0.2ms for T n , since this delay
is mainly propagation delay. In the wide area, for example
from NCSU to locations in California, the RTTs range from
Fig. 5 In-band/out-of-band signaling with hardware signaling,
75ms (California Institute of Technology, CIT) to 89.99ms metro area, µ tx = µ proc
(San Diego State University, SDSU). We choose 35ms and
OOB IBj
45ms for T n , and 25ms for T n (propagation delay). The
reason we allow for both a metro-area and a wide-area setting
is that some service providers may connect two signaling-
enabled switches located across a wide-area with logical links
making them neighbors from a signaling point of view. In the
telephone network, for example, many carriers do interconnect
signaling-enabled switches in California directly via logical
links (for the user-plane interfaces) with switches in New York.
OOB
We choose 1% and 5% for p . As we explained before,
even for in-band signaling, bit- and burst-link errors, buffer
IBj –4
overflows may cause packet loss. For p , we choose 10 Fig. 6 In-band/out-of-band signaling with hardware signaling,
–8
and 10 . wide area, µ tx = µ proc
In the following subsections, we will show the delays under
different scenarios. In each case, we vary the aggregate signal- C. Hardware-accelerated signaling engine, and µ tx « µ proc
ing message arrival rate, from 0.01µ tx to 0.95µ tx , to show
the delays under different loads.

B. Hardware-accelerated signaling engine, and µ tx = µ proc


From Fig. 5, we can observe that in the metro area, when
hardware-accelerated signaling is used and µ tx = µ proc , in-
band signaling always outperforms out-of-band signaling.
Referring to (7), the first two terms are relatively small (in the
order of µs ) when compared to T n , which is in the order of
ms . Thus, the total delay is mainly determined by T n . Since
Fig. 7 In-band/out-of-band signaling with hardware signaling,
IBj OOB metro area, µ tx « µ proc
Tn is less than T n , the in-band solution does better.

According to (7), when p is small, its impact on the total When µ tx « µ proc , for hardware-accelerated signaling,
the first term of (7) is still negligible, but the queueing delay at
the transmitter (i.e., the second term of (7)) becomes compara-
1. http://www.nlanr.net

6
ble to the third term in the metro area setting, where T n is in dominate the total delay.
the hundreds of µs . When ρ is high, the factor In the wide area (Fig. 10), the choice between in-band and
out-of-band signaling really depends on the choices of parame-
( 1 – ρ ⁄ 2 ) ⁄ ( 1 – ρ ) starts increasing to the point where the sec-
ters. When T n = 35ms and p=1%, out-of-band signaling dem-
ond term dominates the total mean delay. In this case the out- onstrates the best performance. But we do note that the RTT
of-band signaling outperforms in-band signaling because µ tx measurements obtained from the NLANR site were made pri-
is a five times higher in the out-of-band case than in the in- marily on Internet2, which is lightly loaded. On the Internet,
band case (see Table 2). In the wide area, T n still dominates RTTs are likely to be significantly higher. In this case, even
with software signaling processors, the network delay may
and hence the lower rate of the in-band channel is compensated dominate, making in-band signaling a better choice.
by the increased T n of the out-of-band solution.
V. CONCLUSIONS
We compared in-band and out-of-band signaling transport
options under assumptions of the switches having hardware-
accelerated signaling engines or software signaling processors.
With hardware signaling engines, if the bandwidth allocated
for the in-band signaling channels is in par with the speeded-up
processing rates, then the main source of delay in sending a
signaling message is the network delay. Using an out-of-band
path through the public Internet could make this component
significant and thus obliterate the gains in call setup delays
made with hardware signaling engines. In this case, fast in-
Fig. 8 In-band/out-of-band signaling with hardware signaling,
band signaling channels are required. In the parameters we
wide area, µ tx « µ proc
considered, this required setting aside one OC1 for each neigh-
bor switch to which there could potentially be a large number
D. Software signaling processor of interfaces, each at a multi-OC1 rate. On the other hand, with
software signaling processors, we found that network delays
matter much less, and hence the cheaper out-of-band public
Internet option is sufficient in the metro area. But in the wide
area, in-band signaling is typically a better option.

ACKNOWLEDGMENT
This work is sponsored by an NSF ANIR grant, 0087487.
This work also benefits from an NSF EIN grant, 0335190, and
an NSF ITR grant, 0312376.

REFERENCES
Fig. 9 In-band/out-of-band signaling with software signaling, metro area
[1] E. Mannie, “Generalized Multi-Protocol Label Switching (GMPLS) Ar-
chitecture,” IETF Internet Draft, May 2003.
[2] L. Berger, “Generalized Multi-Protocol Label Switching (GMPLS) Sig-
naling Resource ReserVation Protocol-Traffic Engineering (RSVP-TE)
Extensions,” IETF RFC 3473, Jan. 2003.
[3] M. Veeraraghavan, X. Zheng, H. Lee, M. Gardner, and W. Feng,
“CHEETAH: Circuit-switched High-speed End-to-End Transport Ar-
cHitecture,” Proc. of Opticomm2003, Dallas, TX, 2003.
[4] H. Wang, M. Veeraraghavan, R. Karri, and T. Li, “A Hardware-Acceler-
ated Implementation of the RSVP-TE Signaling Protocol,” Proc. of
ICC2004, Paris, France, 2004.
[5] L. Berger, D. Gan, et al., “RSVP Refresh Overhead Reduction Exten-
sions,” IETF RFC 2961, April 2001.
[6] V. Paxon and S. Floyd, “Wide-Area Traffic: The Failure of Poisson
Fig. 10 In-band/out-of-band signaling with software signaling, wide area
Modeling,” IEEE/ACM Transactions on Networking, Vol.3, June 1995.
[7] Donald Gross, Carl M. Harris, Fundamentals of Queueing Theory, 3rd
From Fig. 9, we can observe that out-of-band signaling out- edition, Wiley-Interscience, 1998.
performs in-band signaling in the metro area when software
signaling processors are used. This is because the message pro-
cessing delay and emission delays (first two terms of (7)) will

Вам также может понравиться