You are on page 1of 10

Performance analysis of an FDL based optical packet buffer

Dieter Fiems and Herwig Bruneel


SMACS Research Group, Vakgroep TELIN (TW07), Ghent University
St-Pietersnieuwstraat 41, B-9000 Gent, Belgium
Abstract
We evaluate the performance of optical fibre delay line based packet buffers. In particular, we
focus on the packet loss in FDL structures without feedback. Two different FDL selection procedures are investigated analytically using a probability generating functions approach. Further, we
assess the impact of the lengths of the FDL lines on the performance of the buffer and illustrate
our analysis with some numerical results.

Introduction

Future communication networks will have to cope with the ever increasing bandwidth demand related to all kinds of multimedia services over the internet. Optical packet switching (OPS) networks
combine the flexibility of packet switched networks with the bandwidth capacity of optical fibres (see
a.o. [1, 2, 3] and the references therein). An OPS network consists of optical packet switches, interconnected with optical fibres. User data is transmitted in fixed length optical packets consisting of
payload and header information. The payload is switched entirely in the optical domain whereas the
header information is converted to and processed in the electrical domain. Conversion to the electrical
domain is required as there is no optical logic available.
As different packets may contend for the same output, optical buffers are required to mitigate packet
loss. Buffering in the optical domain relies on the use of fibre delay lines (FDLs). That is, the
optical packet is buffered by sending it through an optical fibre. There are two main classes of FDL
structures for buffering: FDL buffers with and without feedback. In the former class, the output of
the FDL structure is fed back to the input if there is contention whereas in the latter class, packets are
dropped if there is contention at the output of the FDL structure.
In the present contribution, we analytically investigate the performance of FDL structures without
feedback for synchronously arriving fixed length packets. We assume that at most one packet can
leave the system at a time. That is, we consider a single server system. Performance assessment of
this type of FDL structures has been considered before. In [4], the performance of an FDL structure
consisting of FDLs of different lengths fed by uncorrelated traffic is investigated. Packets are routed
to the shortest delay line such that there is no contention at the output. The authors note that an exact
solution is not feasible and provide an approximate analysis. An FDL structure fed by uncorrelated
traffic consisting of FDLs with consecutive (single packet length) delays is considered in [5]. Packets
are routed to the first available delay line and may cause contention at the output. Finally, in [6],
performance of an FDL structure fed by correlated traffic is assessed. The FDL structure under investigation consists of FDLs with consecutive multiple packet length delays and as in [4] packets
1

are routed to the shortest available delay line. In this contribution, we compare the routing scheme of
[4, 6] with the scheme of [5] for FDL structures with consecutive multiple packet length delays, fed
by Markov-correlated traffic.
The rest of this contribution is organised as follows. In the next section, we provide a more detailed
description of the FDL structures and the arrival process under consideration. Sections 3 and 4 then
concern the analyses for the different routing schemes. Finally, we illustrate our results by some
numerical examples in section 5 and draw conclusions in section 6.

The FDL Buffer Structure

We consider an FDL structure fed by synchronously arriving traffic. That is, we assume that time is
divided into fixed length intervals or slots and that packets only arrive at (just before) slot boundaries.
The packets have a fixed length and we assume that the slot length corresponds to the transmission
time of a packet. Packets arrive in the FDL structure just before slot boundaries. Furthermore, we
assume that at most 1 packet may leave the FDL structure at the end of a slot. That is, we consider a
single server queueing system.
The structure consists of N delay lines with lengths ng (n = 1 . . . N ). Here denotes the slot
length and g denotes the granularity, an integer constant. Further, a zero length delay is also possible.
In that case, the packet is directly routed to the output.
Arriving packets are routed to available delay lines according to one of the following two routing
schemes:
Scheme NC (no contention): packets are routed to the shortest available delay line such that
there is neither contention at the input nor at the output of the FDL structure. That is, packets
are routed such that during a slot at most a single packet enters a delay line and such that at
the end of a slot at most a single packet leaves the complete delay line structure. The packet is
dropped if such a delay line cannot be found.
Scheme S (simple): packets are routed to the shortest available delay line. That is, packets are
routed such that during a slot at most a single packet enters a delay line. The packet is dropped
if such a delay line cannot be found. Note that this scheme allows contention at the output of
the structure as only one packet can leave the FDL structure during a slot. Excess packets are
dropped.
Note that the S scheme does not require information regarding the state of the FDL structure as
opposed to the NC scheme. Therefore, the S scheme is easier to implement. One easily shows that
a single counter captures the state of the structure with consecutive delays (g = 1). For g > 1, g
counters can capture the state of the system as it decomposes into g virtual FIFO queues (see further).
The state of a more general set of FDLs is more complex to describe.
The number of packet arrivals at the end of the consecutive slots are modeled by means of a discretetime finite-state Markov modulated batch arrival process. At slot boundaries, the arrival process is in
one out of M possible states. Given this state, say state i (i {1 . . . M }), there are k packet arrivals
just before the next slot boundary and the arrival process is in state j at the next slot boundary with
a fixed probability aij (k). Therefore, the arrival process is completely characterised by the matrix
2

A(z) = [Aij (z)]i,j=1...M of (partial) probability generating functions


Aij (z) =

aij (k) z k .

(1)

k=0

Also, for further use, let A(z) denote the probability generating function of the number of arrivals at
the end of a random slot. That is,
A(z) = A(z)eT
(2)

Here eT denotes the M 1 column vector with all elements equal to 1 and denotes the steady-state
probability row vector of the states of the Markovian arrival process. That is, is the non-negative
normalised solution of = A(1).

The packet loss for the NC scheme

We first consider an FDL structure with consecutive single packet length delays (g = 1) and with
routing according to the NC scheme. One easily verifies that the optical buffer with the NC routing
scheme operates as an ordinary FIFO buffer with size (not including the packet in service) equal to
the number of fibre delay lines N . For simplicity, we first focus on an infinite size queueing system
and then approximate the packet loss ratio by means of tail approximations. Kim and Shroff [7] note
that plots of the packet loss ratio in finite capacity queues and of tail probabilities in infinite capacity
queues look very similar. Therefore, they propose following heuristic for the the packet loss ratio
PLR(N ) of a queue with N buffer spaces,
PLR(N ) =

PLR(0)
Pr[U > N ].
Pr[U > 0]

(3)

Here, Pr[U > k] denotes the probability that the queue contents of the infinite capacity queueing
system exceeds k.
The infinite capacity FIFO queueing system with Markov-correlated arrivals has been studied extensively in the past and we may therefore rely on previous results (e.g., [8, 9]). In particular, the tail
probabilities Pr[U > k] are investigated in [9]. The interested reader is referred to this contribution.
As such, we only need to determine the packet loss ratio PLR(0) for the bufferless system. For such
a system, all packets but one that arrive at the end of a random slot are lost. The mean number of
packets that enter the system and leave the system are given by A0 (1) and (1 A(0)) respectively.
Therefore, we obtain following expression for the packet loss ratio:
PLR(0) = 1

1 A(0)
A0 (1)

(4)

Let us now shift focus to the FDL system with consecutive multiple packet length delays (g > 1).
Consider a random slot, say slot n. Arrivals at the end of this slot can only leave the FDL structure
at the end of slots n + 1 + kg for some integer non-negative k due to the assumptions regarding the
lengths of the available delay lines (since it takes kg slots to exit the k-th delay line). Therefore,
these customers may only contend with customers that arrive(d) at the end of slot n + lg for some
integer value l. This observation leads to the conclusion that we may regard the FDL structure under
3

virtual buffer 1

virtual buffer 2

virtual buffer g

Figure 1: The system with g virtual buffers.


queue size: un
arrival state: i

(un 1)+ + aii0


j

i0
aii0 arrivals

slot n + g

slot n

time

departure if un > 0

g 1 transitions

Figure 2: Arrivals in and departures from a virtual queue


consideration as a set of g (virtual) buffers as depicted in figure 1. During the consecutive slots,
arrivals are cyclically routed to the different virtual buffers and the buffers are cyclically served.
Since an arriving packet is routed to the shortest delay line such that there is no contention, each of the
virtual buffers operate as a FIFO queueing system. One should however note that due to the non-unity
granularity, the complete FDL structure does not follow a FIFO queueing discipline. Packets may
leave the FDL structure out of order and the discipline is not work conserving as during some slots no
packets leave the FDL structure while there are packets in it. This is due to the fact that once a packet
is put in an FDL, it cannot be transmitted before it exits the FDL. Or, in other words, FDLs cannot be
accessed randomly.
Now, consider the arrival process into a single virtual buffer. Every g slots, packets are routed to the
virtual buffer under consideration and if there are packets present at the beginning of the slot, a single
packet leaves at the end of the slot. The systems behavior during time is visualised in figure 2. One
now easily verifies that the virtual buffer observed every g slots can be modelled as a FIFO queueing
system with Markov-modulated batch arrivals as considered before (i.e., as considered for the case
4

g = 1). The transition-matrix of the arrival process of this queueing system is given by,

A(z)
= A(z) Ag1 (1).

(5)

The former expression follows from the fact that after slot n (see figure 2), the arrival process goes
through g 1 transitions until slot n + g. There are no arrivals in the virtual buffer under consideration
during these g 1 slots.

As all virtual queues are fed by the same arrival process, one finds that the packet loss ratio of the
system equals the packet loss ratio of a single virtual queue. Summarising we find that adapting the
arrival process by means of expression (5) translates the case of general g into the case g = 1. That
is, we can find the packet loss ratio for general g by means of expressions (3) to (5) and by means of
the results of reference [9].

The packet loss for the S scheme

As for the NC scheme, we first focus on an FDL structure with consecutive delays (g = 1).
Consider a random slot, say slot n, and assume that there are i arrivals at the end of this slot. The
simple routing scheme under consideration then routes the packets such that these arrivals leave the
FDL structure at the end of slots n + 1 to n + min(i, N + 1). Note that packets are lost at the entrance
of the FDL structure if there are more than N + 1 arrivals.
We can now consider the number of packets that leave the FDL structure at the end of slot n + N + 1.
Note that slot n + N + 1 is a random slot as slot n was chosen randomly.
The simple routing scheme implies that not more than one packet that arrived at the end of slot n + j
(0 j < N + 1) can leave the system at end of slot n + N + 1. A packet that arrived at the end of
slot n + j leaves the queue at the end of slot n + N + 1 if there are more than N j arrivals at the
end of slot n + j. Let a(n+j) denote the number of customers that arrive at the end of slot n + j, then
we find that
N
X
1(a(n+j) > N j)
(6)
b(n+N +1) =
j=0

customers leave the system at the end of slot n + N + 1. Here 1(x) denotes the indicator function.

As slot n is chosen arbitrary, the state of the arrival process at the beginning of slot n is described by
the row vector . Recall that the latter is the non-negative normalised solution of = A(1).
(k)

Further, let B (k) (z) = [Bij (z)]i,j=1...M denote the matrix with following (partial) probability generating functions as elements,
k
X
(k)
Bij (z) = z + (1 z)
aij (l).
(7)
l=0

From equations (6) and (7) and taking the Markovian nature of the arrival process into account, we find
that the probability generating function B(z) of the number of departures at the end of slot n + N + 1
is given by,
N
+1
Y
B(z) =
B (N k+1) (z) eT
(8)
k=1

Here eT denotes the column vector with all elements equal to 1. Recall that slot n+N +1 is a random
slot as k was chosen randomly. Therefore B(z) also denotes the probability generating function of
the number of packets that try to leave the FDL structure at the end of a random slot. As at most
one packet can leave the system at the end of a slot, one easily verifies that (1 B(0)) equals the
mean number of departures from the system at the end of a random slot. Therefore, we find following
expression for the packet loss ratio,
PLR(N ) = 1

1 B(0)
.
A0 (1)

(9)

We can now again focus on an FDL structure with longer delays (g > 1). As for the NC-scheme,
there is no contention between packets that arrive in slots n and l for (n l) mod g 6= 0. Therefore,
the system also operates as one with g virtual queues (see section 3). We can focus on the packet loss
ratio in a single virtual queue and note that the latter also equals the packet loss ratio of the complete

system. The arrival process in a single virtual queue is again described by the matrix A(z)
as given in
expression (5) and the S-scheme is applied in each virtual queue. Therefore, adaptation of the arrival
process again translates the case of general granularity g into the case g = 1 as investigated before
(see expressions (7) to (9)).

Numerical examples

To evaluate the FDL structure numerically, we assume that the arriving packets come from the compound traffic of 2 types of 2-state Markovian sources. During the consecutive slots, each source
alternates between an off state and an on state. A source generates a single packet if it is in the onstate and no packets if it is in the off-state. There are ni sources of type i (i = 1, 2). The transition
probabilities from the on-state to the off-state and from the off-state to the on-state for a source of
type i (i = 1, 2) are given by (1 i ) and (1 i ) respectively. The compound source is therefore
completely characterised by the tuple (1 , 1 , n1 , 2 , 2 , n2 ). Alternatively, we may characterise the
source by the tuple (1 , K1 , n1 , 2 , K2 , n2 ). Here i and Ki (i = 1, 2) denote the arrival rate and
burstiness factor of a single source:
i =
Ki =

1 i
,
2 i i
1
,
2 i i

(10)
(11)

for i = 1, 2. The burstiness factor Ki is a measure for the absolute lengths of the on- and off-periods
and takes values between max(i , 1 i ) and infinity. For Ki = 1, the source turns into a Bernoulli
source and therefore the generated traffic becomes uncorrelated.
Figure 3 depicts the packet loss ratio versus the number of fibre delay lines for both the NC routing
scheme (a) and the S routing scheme (b). For each of these schemes, different granularities g =
{1, 2, 4} are assumed as depicted. In the case of the NC scheme, we also depict some simulation
results as our results are only approximate. Arrivals come from a superposition of n 1 = 3 correlated
sources (1 = 0.1, K1 = 20) and n2 = 3 uncorrelated sources (2 = 0.1, K2 = 1) for all plots. As
expected, the NC scheme clearly outperforms the S scheme. Further, an increase of the number of
fibre delay lines implies a decrease of the packet loss ratio for both schemes. For the NC scheme this
6

0.22

0.215
0.21
0.1

0.205
0.2
PLR

PLR

g=1
0.01

0.195

g=2

g=1

0.19

g=2

0.001

0.185
g=4

0.18
0.0001

10

15

20

25

30

35

40

number of fibre delay lines N

(a) NC routing scheme

0.175

g=4

4
6
number of fibre delay lines N

10

(b) S routing scheme

Figure 3: Packet loss ratio vs. the number of Fibre Delay lines

decrease is exponential. This is not unexpected as the FDL structure with the NC scheme corresponds
to a FIFO queue for g = 1 and to a set of FIFO queues for g > 1. Also for the S scheme we observe
that an increase of the number of FDLs implies a decrease of the packet loss ratio. However, the
packet loss ratio does not decrease further for relatively low values of the number of FDLs. This
comes from the fact that packets are always routed to the shortest delay lines. The (additional) longer
delay lines are only used if there arrives a large batch at the end of a slot. We can further note that
increasing the granularity decreases the packet loss ratio. For higher granularity, the correlation in the
arrival process to a virtual queue the packet loss ratio of the complete system equals the packet loss
ratio of a virtual queue becomes less and less correlated which implies a decrease of the packet loss
ratio in that virtual queue.
Figure 4 depicts the packet loss ratio versus the arrival load for both the NC routing scheme (a) and
the S routing scheme (b). For each of these schemes, different granularities g = {1, 2, 4, 8} are
assumed as depicted. The arrival load comes from the superposition of 4 correlated (K 1 = 20) and 4
uncorrelated (K2 = 1) sources. 20% of the traffic comes from the uncorrelated sources and there are
7 fibre delay lines. As expected, for both schemes, an increase of the arrival load implies an increase
of the packet loss ratio. Again, we can observe that the NC scheme always outperforms the S scheme.
However, for higher load, the packet loss ratios of the S scheme is not this much worse than the packet
loss ratio of the NC scheme.
In both figures 5 and 6 we investigate the influence of correlation in the arrival process on the packet
loss ratio for the NC scheme (a) and the S scheme (b). In figure 5 we assume that arrivals come from
the superposition of 10 sources. Sources are either correlated ( 1 = 0.08, K1 = 20) or uncorrelated
(2 = 0.08, K2 = 1) and we vary the number of sources n1 that are correlated. Different values
of the granularity g = {1, 2, 4} are assumed as depicted. Clearly, more correlation in the arrival
process implies more loss. Also, we see that the NC scheme outperforms the S scheme. Further, the
packet loss ratio does not depend on the granularity if there is no correlation in the arrival process
7

0.3

g=1
g=2
g=4
g=8

0.25

0.25

0.2

0.2

PLR

PLR

0.3

0.15

0.15

0.1

0.1

0.05

0.05

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

g=1
g=2
g=4
g=8

0.1

0.2

0.3

0.4

arrival load

0.5

0.6

0.7

0.8

0.9

arrival load

(a) NC routing scheme

(b) S routing scheme

Figure 4: Packet loss ratio vs. the arrival load

0.25

0.36

g=1
g=2
g=4

g=1
g=2
g=4

0.34
0.2

0.32

0.15
PLR

PLR

0.3

0.28

0.1

0.26
0.05

0.24

0.22

10

Number of correlated sources

Number of correlated sources

(a) NC routing scheme

(b) S routing scheme

Figure 5: Packet loss ratio vs. the number of correlates sources

10

0.2

0.3

g=1
g=2
g=4

0.18

g=1
g=2
g=4

0.28
0.16

0.26

0.14

0.24
PLR

PLR

0.12
0.1

0.22

0.08
0.06

0.2

0.04

0.18
0.02
0

0.16

10

10

Burstiness factor K

Burstiness factor K

(a) NC routing scheme

(b) S routing scheme

Figure 6: Packet loss ratio vs. the burstiness factor K

(n1 = 0). In this particular case, larger g does not reduce the correlation in the arrival process of a
virtual queue (and therefore also reduces the packet loss ratio) as there is no such correlation. For
higher n1 , additional granularity reduces the packet loss.
Finally, in figure 6 the packet loss ratio is depicted versus the burstiness factor K of the arrival process.
Here we assumed that there are 8 homogeneous sources and that the total arrival load equals 80%.
Again, there are 7 FDLs available to mitigate packet loss. Increasing correlation again implies more
loss. As in figure 5, we observe that the packet loss ratio does not depend on g for uncorrelated traffic
(K = 1). For larger K, an increase of the granularity implies less loss.

Conclusions

We considered the performance analysis of an FDL structure with granularity. Two different routing
schemes were investigated. The simple NC scheme clearly outperformed the S scheme as expected at
the cost of more complexity. The gain of the more complex NC scheme is however limited for high
load.
Further, we showed that granularity can reduce packet loss considerably in comparison with a structure
with the same amount of FDLs but without granularity. This is a result from the fact that granularity
breaks down correlation in the arrival process.

References
[1] R. Tucker and W. Zhong. Photonic packet switching. IEICE Transactions on Communications,
E82-B(2):254264, February 1999.
9

[2] S. Yao, B. Mukherjee, and S. Dixit. Advances in photonic packet switching: An overview. IEEE
Communications Magazine, 38(2):8494, 2000.
[3] L. Xu, H. Perros, and G. Rouskas. Techniques for optical packet switching and optical burst
switching. IEEE Communications Magazine, pages 136142, 2001.
[4] P. Cadro, A. Gravey, and C. Guillemot. Performance evaluation of an optical transparent
packet switch. In Proceedings of the COST 257 First Management Committee Meeting, page
COST257TD(97)12, Leidschendam, 22 23 Januari 1997.
[5] J. Walraevens, S. Wittevrongel, and H. Bruneel. Calculation of the packet loss in optical packet
switches: an analytic technique. International Journal of Electronics and Communications
57(4):270276, 2003.
(AEU),
[6] D. Fiems, J. Walraevens, and H. Bruneel. Discrete time queueing analysis of a fibre delay line
structure with granularity. In The 8th IFIP Working Conference on Optical Network Design &
Modelling (ONDM 2004), pages 5769, Ghent, Belgium, February 2004.
[7] H. Kim and N. Schroff. Loss probability calculations and asymptotic analysis for finite buffer
multiplexers. IEEE/ACM Transactions on Networking, 9(6):755768, 2001.
[8] S. Li. A general solution technique for discrete queueing analysis of multimedia traffic on ATM.
IEEE Transactions on Communications, 39(7):1115 1132, 1991.
[9] B. Steyaert, Y. Xiong, and H. Bruneel. An efficient solution technique for discrete-time queues
fed by heterogeneous traffic. International Journal of Communication Systems, 10:7386, 1997.

10

You might also like