VC Merging in ATM-LSRs
Ahmad Rostami Seyed Mostafa Safavi
Iran Telecommunication Research Center (ITRC)
Amir-Kabir University of Technology
arostami@itrc.ac.ir
Abstract. MPLS is a new technology that has been
proposed by the IETF community and is one of the most
appropriate techniques to provide connectionless IP
services over powerful ATM switches. A MPLS switch,
which is implemented based on an ATM switch, is called ATM-
LSR. VC merging is a critical task for ATM-LSRs because
without supporting VC merging each ATM-LSR has to
manage O(n2) VC values in case of full mesh connectivity.
In this paper, we analyze VC merging in ATM-LSRs,
which have the ability to support Differentiated Services.
We analyze partial VC merging under different service
disciplines and derive some equations for the mean
waiting time of output queuing in ATM-LSRs.
1. Introduction
During recent years, Internet growth in terms of
increasing number of the users and the demands for more
and more bandwidth as well as new emerging services
which need guaranteed QoS has caused an increase in the
gap between transmission capacity of media (such as fiber
optic) and switching capacities in switch-routers.
Therefore, specialists in this field have begun deep
researches in order to keep pace of the transmission speed.
Result of researches in this area has concluded to
emerge a new technology called MPLS, which was
introduced by the Internet Engineering Task Force (IETF)
in 1997.
One of the most interesting aspects of MPLS is that it
is independent of any specific layer-2 (link layer) and
layer-3 (network layer) protocols. Nevertheless, it seems
that using the IP protocol as the network layer protocol
and using ATM as layer-2 protocol for MPLS switches is
the best way of capturing power of both IP protocol
and ATM switching. In MPLS once a data packet enters
the MPLS domain, it is processed by the edge LSR (Label
Switch Router) and a label is assigned to it which
introduces FEC (Forwarding Equivalence Class) of the
packet. Then the labeled packet is switched in the network
only based on its label without any further processing.
When ATM is used as switching technology, the label
is encapsulated in the VPI and/or VCI fields of the header
of the ATM cell and VCs are formed by the MPLS control
protocol.
An important problem in this case is VC merging
technique, which is a critical task to ATM-LSRs in term
of scalability. In VC merging, all incoming cells
belonging to the same FEC receive the same outgoing VC
values (VPI/VCI). That is, all incoming cells to a specific
ATM-LSR, which may have different VC values but have
the same destination network and QoS receive the same
outgoing VC values. As stated above, VC merging has an
important role in scalability of MPLS. If we assume not to
use the VC merging and have n switch-routers which are
connected in full mesh connectivity then each of them has
to manage O(n2) VC values. Obviously we will be faced
with many problems when network nodes increase. On the
other hand, if we use VC merging technique, we can
prevent managing such a big amount of VC values in a
LSR. When we implement MPLS over ATM we have to
use AAL5 encapsulation in which the ordering of the cells
belonging to a specific frame is not conveyed along the
cells to destination, so the ordering of the cells belonging
to a specific frame must be saved in the network. When
we use VC merging we have to implement some
techniques to prevent interleaving cells of different
frames. These considerations impose some problems in
terms of additional delays and additional buffers in LSRs.
That is, by implementing VC merging, we have to store
all cells belonging to a specific frame in a separate buffer
until the last cell of that frame is received (the last cell of a
frame is detected by checking a bit in the PTI (Payload
Type Identifier) field of the header of an ATM cell). After
the last cell of the frame is received we can transfer all
cells continuously without permission to any other cell,
belonging to another frame, to be interleaved by cells of
the frame.
As stated in [1], we may have two kinds of VC
merging. If all transit traffic is best effort, we can
implement full merging, in which all incoming cells with
the same destination network are mapped to the same VC
and so receive the same VC values. On the other hand, if
we have different kinds of transit traffic, we may
implement partial merging, in which all incoming cells
with the same destination network receive the same
 outgoing VC.
In contrast with full merging, in partial merging we
have N different output queues, each of them dedicated to
a particular QoS class. In [1], all performance issues about
VC merging such as performance issues related to
reassembly buffers and the output buffer in full merging
have been studied. In this paper, we study the partial mode
VC merging under two different service disciplines.
2. Architecture of a Partial VC Merger
Arrivals to each of the buffers are batch arrivals and we
model the arrival process to each buffer as a Poisson
process. We define λi ; (i=1, …,6) (frame/cell-time) as the
arrival rate of the class- i frame, ni as the mean number
of the cells in a class-i frame, ρi = λi ni as the mean
offered load to the server due to class-i traffic and
6
ρ = ∑k =1
ρ k as the total offered load to the server.

In this section we introduce structure of the partial VC

merging. Firstly, we assume that the switch is output Q1 (EF) Scheduling
buffered and thus speedup of the switch is N. According
to Fig. 1 in each output port of the switch there are some Q2 (AF1)
reassembly buffers, which save the cells belonging to a Output Link
specific frame until the last cell of the frame is received. Q3 (AF2)
Then all cells of a completed frame are continuously Server
transferred to the merger. If the VC values are unique to
the entire switch rather than to the port then an RB would Q4 (AF3)
correspond to an incoming VC [1]. In the VC merger, VC
values of the received cells are translated to the outgoing Q5 (AF4)
VC values. In this point, as stated above all cells
belonging to the same destination network and QoS are Q6 (BE)
mapped to the same VC values. Then, after determining
the QoS class of the frame, all cells are transferred to the Fig. 2. Structure of Output Queues
corresponding output buffer continuously.

Reassembly Buffers Output Buffers 3.1. Prioritized Service Discipline

Frame 1 1
2 are served according to a prioritized service discipline.
Frame 2
Merging
Frame 3
6
Fig. 1. Partial VC Merging Structure
Regarding to the fact that our switch supports
Differentiated services model, we assume that there are
six different output buffers in the system. First queue is
dedicated to EF (Expedited Forwarding) traffic, Second
through fifth queues are dedicated to AF1 through AF4
(Assured Forwarding) traffics respectively, and the last
queue is for BE (Best Effort) traffic.
3. Performance Analysis
In this section we analyze output buffers and derive
some equations for calculating the mean waiting time for
them under two different service disciplines. According to
Fig. 2 output buffers consist of 6 independent buffers.
Here, we analyze the situation in which output queues
That is, Q1 has strict priority over the rest of the queues;
Q2 has strict priority over the rest of the queues except for
Q1 and so on.
To calculate the mean waiting time for all queues under
this service discipline we can use eq. 1.
Wi = Ri + Li + Mi i=1,…,6 (cell times) (1)
Where Li is the average length of the Qi in the steady
state condition and can be substituted by Little’s formula.
To calculate the mean waiting time of Q1 we can solve
eq. 1 for W 1 :
R1 + M 1
W1 = (cell times) (2)
1 − ρ1
Where R1 is the mean residual time of completion of
service for the queue that is being served when sample
cell enters to Q1. In our system R1 is equal to ρ 2 .
Also, M1 is mean number of the cells which are in front of
the tagged cell in a same frame and can be calculated
using eq.3.

2 2
ni 1
Mi = − (cells)
2ni 2
2 6
Where ni and ni are first and second moments of
number of the cells in a frame of type i respectively.
To calculate the mean waiting time for Q2 through Q6
we can write Ri according to eq. 4.
ρ i −1
Ri = + ∑ (Lk + ρkWi ) i=2,…,6 (cell times) (4)
2 k =1
Substituting Little’s formula and eq. 4 in eq.1 and
solving it for Wi , we can write:
i −1
ρ + ∑ ρkWk + Mi
2 being served by the server (Fig.3-a), which we call this
Wi = k =1
i
i=2,…,6
1 − ∑ ρi
k =1
(cell times)
3.2. Mixed Service Discipline
Now, we analyze the output buffers under mixed
service discipline, which is defined as follows. We assume
that Q1 (EF traffic) has strict priority over the other
queues. Also, other five queues receive service based on
the WRR scheduling algorithm, in which a weight is
dedicated to each queue and each queue receives service
in a round robin order. The weight of each queue is
proportional to a certain fraction of the output link
capacity and each queue is served in its turn based on its
weight. Hence, we define the weight vector m=(m2, m3,
m4, m5, m6) for the system. It should be mentioned that
dedicated priority to Q1 is not preemptive. That is, when a
batch arrives in Q1 and server is serving another queue,
arrived cells have to wait until completion of service
phase of the queue which is being served. Regarding to
strict priority of Q1, the mean waiting time for this queue
again can be calculated using eq.2 where R1 in this
situation can be calculated using the mean residual life
time formula [2]:
W m=(m2, m3, m4, m5, m6) (as defined previously) as the
R1 = (1 + C1 2 ) (cell times)
2 Also, the mean vacation period for each vacation system
In eq.6 W is the mean life time and C1 2 is squared
coefficient of variation ( C1 = σW
2
) . In this service
(3) W
discipline, we can write eq.7 for the mean life time.
W = ρ 1 + ∑ ρkmk (cell times) (7)
k =2
In order to derive equations for the mean waiting time
of Q2 to Q6 we are involved with a cyclic type multi-
queue system with the limited service discipline [2],
which is interrupted by a high priority queue (Q1). In
general, there is not any solution for multi-queue systems
under limited service discipline. In order to overcome this
problem we use a model called “Vacation Model”. A
vacation model is composed of a single queue and a single
server. Before proceeding, we present some definitions
about the vacation models. As depicted in Fig.3, a queue
in a vacation model may have two phases. Either it is
phase as the service phase or the server has stopped
serving the queue and has gone to the vacation (Fig.3-b),
which we call this phase as the vacation phase. Whenever
the server goes to the vacation, it remains in the vacation
(5)
for duration of V (cell time), which we call this duration as
the vacation period. Also, when it comes back from the
vacation, it serves the queue for the duration of m (cell
time), that we call it as the service period. In addition, we
consider summation of the service period and the vacation
period (m+V) as the total period of the vacation system.
Queue Server
(a)
Queue Server
(b)
Fig. 3. Modeling of a Vacation System
Now, by using the above mentioned model we can
begin deriving equations for the mean waiting time of Q2
to Q6. We model our system by considering a separated
vacation model for each of our five queues. We use
(6) service period for each of the assumed vacation models.
can be calculated using eq. 8.
 6 over the rest of the queues. The additional delay equals to
Vi = ∑ρ k .mk + L1 i=2,…,6 (cell times) (8) the average length of Q1 in steady state condition plus
k = 2, k ≠ i mean number of the cells which enter to Q1 during the
time arrival when the tagged cell enters to the system and
Now, we have five separated vacation models with when it departs the system.
service periods and mean vacation periods as defined Finally, we can write the mean waiting time for Q2 to
above. Whenever a tagged cell enters to the Qi ( i=2,…,6), Q6 as eq.12.
it may finds the queue either in the vacation phase or in
the service phase. We will analyze these two phase Mi
separately. The probability that a tagged cell finds the P v , i.( R i + (V i + m i ))
server in a vacation phase when it enters into the queue Wi ≅ mi
can be calculated as follow. ρi
(1 − .( V i + m i ) − ρ 1 )
mi
6 mi
∑ρ
k = 2,k ≠i
k .mk + L1 mi
(1 − Pv, i )( ρi. +
Mi − ρi.
2 (Vi + mi ) + V i ) + L1
Pv , i = 6
i=2,…,6 (9) 2 mi
+
∑ ρk .mk + L1
k =2
(1 −
ρi
.(Vi + mi ) − ρ1)
mi
Clearly, the probability that the tagged cell finds the
queue in the vacation phase when it enters to the queue is i=2,…,6 (cell time) (12)
equal to (1- Pv , i ).
If the tagged cell enters in the vacation phase the mean
waiting time for Q2 to Q6 approximately equals to mean
number of cell times which the tagged cell has to wait 4. Numerical Results
until the server comes back from vacation plus mean
number of total periods which, the tagged cell has to wait In this section we use two different traffic scenarios to
before receiving service. Thus, the mean waiting time in analyze performance of the system and compare the effect
this situation can be written as eq.10. of using different service disciplines in the system. In our
study we use the equations which were derived in sections
3.1 and 3.2.
Wi ≅ Rv , i +
Mi + Li
mi
(mi + Vi ) i=2,…,6 Firstly, we assume that mean number of the cells in
each frame is 10 cells. Also we assume that the weight
(cell times) (10) vector which is used by WRR scheduler is m=(11,9,7,5,7).
In the first scenario 10, 15, 15, 15, 15 and 30 percents of
Where, Rv , i is the mean residual time of the Vacation total offered load belong to EF, AF1, AF2, AF3, AF4 and
phase of Qi and can be easily calculated using the mean BE traffic respectively. In the second scenario, share of
residual time formula (similar to eq.6). mentioned traffic classes change to 5, 5, 5, 10, 10 and 65.
If the tagged cell enters to the system in the service Fig.4 through Fig.10 show results of using derived
phase then the mean waiting time approximately equals to equations for EF, AF and BE traffic classes.
number of the cells, which are served in current service As depicted in Fig.4, the mean waiting time for EF
phase plus mean number of total periods, which the traffic class is small and completely bounded in
tagged cell has to wait before receiving service. Thus, the comparison with other traffic classes. Also, the mean
mean waiting time in this situation can be written as waiting time under prioritized service discipline is smaller
eq.11. than that of mixed service discipline. The reason of such
situation is that the mean residual time of completion of
mi service for Q1 (R1 in eq.2) in the prioritized service
Mi + Li − ρi discipline equals to ρ 2 ; however, according to eq.6,
mi
Wi ≅ ρi + Vi +
2 mi
(
2 mi + Vi ) that of the mixed service discipline is obviously greater
than ρ 2 .
i=2,…,6 (cell times) (11)
As we expected, the mean waiting time increases when
Whether the tagged cell enters to the system in the we move from EF traffic class toward BE traffic class.
vacation phase or in the service phase, it has to tolerate an According to Fig.5 through Fig.8 the mean waiting time of
additional delay due to the EF traffic which has priority AF1, AF2 and AF3 traffic classes in prioritized service
discipline is smaller than that in mixed service discipline
under the same traffic scenario. In AF4 traffic, the mean
waiting time under traffic scenario II in prioritized service Scenario I
discipline is smaller than that in mixed service discipline. Scenario II
There is the same situation scenario I whenever the total
offered load is less than about 90%. For total offered load Mixed
above 90% the mean waiting time for the prioritized discipline
service discipline exceeds that in mixed service discipline.
The reason is that in scenario I 55% of the total offered
load has priority over the AF4 traffic class and this leads
the mean waiting time to be increased. However, in Prioritized
discipline
scenario II only 25% of the total offered load has priority
over the AF4 traffic class.
The mean waiting time for BE traffic class (Fig.9 and
Fig.10) under prioritized service discipline is dramatically
increased in comparison with the mean waiting time under Fig. 5. Mean Waiting Time for Q2 (AF1)
the mixed service discipline. That is why we depicted
curves of different service disciplines separately. As
depicted in Fig.9, the mean waiting time for the BE traffic
class is extremely high in comparison with other classes,
because this class has the lowest priority despite the fact Scenario I
Scenario II
that a big fraction of the total offered load belongs to it. Mixed
The situation is exacerbated when the total offered load discipline
approaches 100%. Also, in prioritized service discipline
the mean waiting time for BE traffic class in scenario II is
less than that in scenario I, because in scenario II the
volume of total load which belongs to BE traffic is greater
than that in scenario I. Prioritized
Fig.10 shows the mean waiting time for BE traffic class discipline
under mixed service discipline. Here, the mean waiting
time is meaningfully less than that of the prioritized
service discipline, because in this situation traffic of type
BE has access to a certain amount of the output link
capacity; however, in prioritized service discipline Q6 Fig. 6. Mean Waiting Time for Q3 (AF2)
(BE class) is served whenever other queues are empty.

Scenario I
Scenario I Scenario II
Scenario II
Mixed
Mixed
discipline
discipline

Prioritized
discipline Prioritized
discipline

Fig. 7. Mean Waiting Time for Q4 (AF3)

Fig. 4. Mean Waiting Time for Q1 (EF)
 Services. Also, we derived some equations to calculate the
Scenario I mean waiting time in output ports of such ATM-LSRs for
Scenario II two different service disciplines. In addition, we used
derived equations under two different traffic scenarios and
compared the results. We recognized that the mean
Mixed waiting time for EF and AF classes in prioritized service
discipline discipline is less than that in mixed service discipline
(Priority queuing and WRR scheduling). However, the
Prioritized mean waiting time for BE traffic in prioritized service
discipline discipline is meaningfully high.
We conclude that if only the delays of EF and AF
classes are important and the delay of BE traffic is not too
important, it is better to use prioritized service discipline.

Fig. 8. Mean Waiting Time for Q5 (AF4) 6. References

[1] I. Widjaja, A. I. Elwalid, ”Performance Issues in VC-Merge
Capable Switches for Multiprotocol Label Switching”,
IEEE JSAC., Vol. 17., No.6, 1999.
Scenario I
Scenario II [2] H. Akimaru, K. Kawashima, “Teletraffic ,Theory and
Applications”, Springer-Verlag, 1993.

[3] B.Davie, Y.Rekhter, “MPLS Technology and

Applications” , MORGAN KAUFMANN, 2000.

[4] Tedijianto, “ NonExhaustive Policies in Polling

Systems and Vacation Models: Qualitative and
Approximate Approaches”, Ph.D. thesis, University of
Maryland, 1990.

[5] C. Dou, C. Lin, S. Wang, K. Leu, “Performance

Analysis of Packet-Level Scheduling in an IP-over-ATM
Network with QoS Control”, IEICE Trans. Communications,
Fig. 9. Mean Waiting Time for Q6 (BE) under
Vol.E83-B, No.7 , 2000.
Prioritized service discipline
[6] B.Davie, J.Lawrence, et. al., “MPLS using LDP and VC
Switching”, IETF, RFC 3035, January, 2001.
Scenario I
Scenario II [7] E.Rosen, A.Viswanathan, R.Callon, “Multiprotocol Label
Switching Architecture”, IETF, RFC 3031, 2001.

Fig. 10. Mean Waiting Time for Q6 (BE) under

Mixed service discipline

5. Conclusions

In this paper we have studied the partial mode VC

merging in ATM-LSRs, which support Differentiated