Performance of Switching Fabrics Used for the Scalable Router
Zbigniew Hulicki
Department of Telecommunication, AGH University of Science and Technology, ul. Czarnowiejska 78, Krakow, Poland
Keywords: Switching Fabric, Multistage Interconnection Network (MIN), Scalable High-Performance Router.
Abstract: The intention of this paper is to examine the performance aspects of switching fabrics which can be used for
scalable high-performance routers. Topology and capabilities of switching fabrics are discussed, followed
by an examination of their performance measures in diverse scenarios for possible operation of switching
elements (SEs). It has been shown that the switching fabric based on the cube type multistage
interconnection network (MIN) outperforms those based on the PM2I type MIN under both any traffic
scenario and SE operation. The simulation results presented should be helpful in predicting the router
performance before actual fabrication of the switching fabric.
1 INTRODUCTION
Routers and switches are viewed as the most critical
parts of the current communication infrastructure
and will be also needed to provide fast and efficient
communication in next-generation networks
.
Recently there have been various efforts to design an
efficient optical switch for use in a high-speed router
(
Chao, 2007). Besides, there is an immense interest in
designing a simple and high performance switch
which would satisfy the demands of an entirely new
scenario for emerging broadband services.
In the past few decades, a number of switch
fabrics have been proposed in literature and used in
practical implementations to interconnect key
components in routers, such as routing engines and
line cards (Lien, 2010). Many of the proposed
switch designs have been based on multistage
interconnection networks (MINs). Various solutions,
i.e. single or multiple panel (replicated), can be used
in design of MIN based switches (Tzeng, 2004).
The purpose of this paper is to examine the
issues dealing with the performance of single plane
MINs used as switch fabrics for scalable high-
performance routers and/or optical cross-connects
OXCs. Topology and capabilities of switching fabric
will be discussed first, followed by a specification of
the switch model. Then, the performance and
dependability of MIN based switch fabrics will be
examined, taking into account diverse scenarios for
possible operation of SEs. Lastly, there are some
conclusions and remarks regarding the trade-offs
between switch characteristics.
2 TOPOLOGY OF SWITCHING
FABRIC
In selecting the architecture of switching fabric, four
design decisions can be identified (
Chao, 2007):
operation mode, control strategy, switching method,
and topology of MIN. The topology of a switching
network is a key factor in determining a suitable
fabric architecture (Skalis, 2010). MINs proposed
for the switch fabrics are usually constructed using
small crossbar switches organized in stages and may
have a uniform or non-uniform connection pattern
between stages.
SEs can be buffered or non-
buffered, but the use of self-routing MINs are
favored because small delays can be achieved
(Rongsen, 2007). Generally, the topologies of MINs
tend to be regular and can be grouped into two
categories: static or dynamic.
The static MINs are
simple to build, but fail even in the presence of
a single fault. On the other hand, a dynamic MIN is
able to reroute the data through a fault free path if
the regular path is faulty or busy. Because the
reliable operation of a MIN is an important factor in
overall router performance, it is important to design
a switching network that combines full connection
capability – in spite of faults – with a slightly lower
performance.
A number of diverse fabric solutions which offer
redundant paths between the input and output ports
567
Hulicki Z..
Performance of Switching Fabrics Used for the Scalable Router.
DOI: 10.5220/0005033005670571
In Proceedings of the 4th International Conference on Simulation and Modeling Methodologies, Technologies and Applications (SIMULTECH-2014),
pages 567-571
ISBN: 978-989-758-038-3
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
have been proposed to provide fast, efficient and
reliable communications at a reasonable cost (cf.
Chao, 2007). However, no single solution network is
generally preferred because the cost-effectiveness of
a particular design varies with such factors as its
application, the required speed of data transfer, the
actual hardware implementation of the switching
network, the number of input/output ports, and the
construction cost.
As far as MINs are concerned, they can be built
using a variety of structures. Two significant
examples of topological structures based on the
single plane design are switching networks based on
the cube
i
and PM2I interconnection functions (Lien,
2010). Therefore, there is an open question about
the performance vs.-reliability trade-off that
a multiple-path structure of the PM2I network
might
offer, for various distributed control algorithms and
decision rules, when compared to a cube-type
network.
3 THE SWITCH MODEL
To evaluate the properties of a fabric based on the
given switching network (MIN), the following
assumptions are made about the operation and the
environment of the interconnection network: the
single plane design of switching fabric is taken into
account; because, in this contribution, we are
focused primarily on the performance, the slotted
traffic source model is used with a uniform random
distribution of packet destinations; the switching
network is operated synchronously, meaning that the
packets are transmitted only at the beginning of
a time slot given by the packet clock and each input
link is offered the same traffic load; the buffering is
external to the switch fabric, i.e. non-buffered switch
fabrics formed from non-buffered SEs are under
consideration. Therefore, the queuing effects are
unaccounted in this model. All output ports perform
the function of a perfect sync and a conflict is said to
occur if more than one packet arrives at the same
output link at the same time.
Two classes of packet switched MINs, one with
redundant interconnection paths and one without are
under consideration (i.e. the cube-type class of MINs
known as unique-path networks and the PM2I class
of MINs which have multiple paths between a given
network input-output pair). It is clear that faults in
the switching network often result in degraded
performance of a fabric. Because failures occur at
random and the combinations of the type, number
and location of faults can also vary, the influence of
failures on the performance of switching fabric is
usually examined by simulation. These aspects have
been already investigated (Hulicki, 2013) and hence
they are not included in the model of switching
fabric used in this contribution.
4 PERFORMANCE EVALUATION
Although recently a new (analytic) method for the
performance analysis of multistage switching fabrics
has been proposed (Hulicki, 2013a), in this paper,
the performance of switching fabric, based on both
of the aforementioned classes of packet switched
MINs, has been studied using simulated
experiments. A discrete time event-driven simulator
has been designed to carry out simulations. Such
a tool enables one to simulate the real operation of
switch fabric as well as to estimate its performance
measures. The simulator uses the switch model and
the most commonly used performance measures. It
enables performance evaluation of the fabric under
different traffic and/or interconnection patterns,
including both uniform and non-uniform traffic
patterns. Moreover, in order to stop packets from
entering the switching fabric once its resources are
exhausted, the switching network of a router often
employs a backpressure mechanism (Lien, 2010).
Therefore, except for non-buffered MINs, the
simulator also allows for a modification to the
switching fabric, i.e. an implementation of buffering
(introduced at the input/output of the switching
network or to the SEs) or using different types of
SEs (that allow for diverse switching functions), as
well as an evaluation of the fabric’s dependability
under diverse fault models, including combinations
of the type, number and location of faults.
As it has been already mentioned, this
contribution presents only selected results from the
simulation experiments because the results point to
a similar trend and the utilization of fabric,
efficiency of switching and packet loss (or packet
drop rate) can serve as good indicators for the
performance of switching fabric. The simulation
results have been shown and compared in a few
figures (Figs. 2, 3, 4, 5, 6 and 7). Each data value
given in these figures is the result after n
r
= 100,000
clock cycles in the simulation (cf. Fig.1), where this
number of cycles is found to yield steady-state
outcomes (95% conf. lev.). If the number of cycles
is less than 10,000 , the steady-state is not reached,
so the simulation results cannot be valid (cf. lowest
curve in Fig.1).
SIMULTECH2014-4thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
568
average packet delay
9
5
%
conf.
lev
. .
unif.
traffic
.
distr.
in
;
out
N
=
6
4
g
x
offered
load
Figure 1: Comparison of packet delay for various
simulation clock cycles.
total nr of switched packets
j
= 2
j
= 3
j
= 4
j
=1
j
= 0
offered
load
(
n
-
j
)
N = 32
18,0
0,0
2,0
4,0
6,0
8,0
10,0
12,0
14,0
16,0
0,0 0,2
0,4
0,6
0,8
1,0
g
x
Figure 2: Total number of switched packets vs. traffic load.
The total number W
(n-j)
of switched packets at
the outputs of the given ((n-j)–th) fabric stage versus
traffic load offered to the input ports is demonstrated
in Fig. 2. It is observed that increasing traffic load
results in nonlinear increase of W
(n-j)
, however
along with that increase the inflexion of
characteristic is also moved towards greater values
of both parameters. The reason for that phenomenon
is the cumulative effect of events occurring during
a packet cross over the switching fabric.
Influence of SE operation mode on the
performance of the cube type (MIN) fabric is shown
in Fig. 3. One can see that implementation of
switching function with the broadcast capability
decreases fabric performance, even though it could
potentially increase switching capabilities of SE.
Implementation of such switching function increases
packet drop rate at all stages in the fabric what
deteriorates its performance measures.
As it has been already mentioned, the buffering
is external to the switch fabric, i.e. non-buffered
switch fabrics formed from non-buffered SEs are
under consideration in this performance study.
Hence, it was interesting to investigate the queuing
effects at the inputs of switching fabric for its
total nr of switched packets
1
3
5
2
4
0
0,0 0,3 0,5 0,8 1,0
SE with the broadcast
capability
j
= 0
j
= 0;1
j
= 0
j
= 0;1
j
= 0;1;2
SE without the broadcast
capability
j = 0;1 denotes:
after crossing two stages
W
j-n
g
x
offered
load
Figure 3: Influence of SE operation mode on the fabric
performance.
=32
N=64
Q
_
input queue length
g
x
0.0 0.2 0.4 0.6 0 . 8 1 . 0
offered load
10
30
50
0
20
40
N=4
N=8
N=16
N
Figure 4: Queuing of packets at the input of switching
fabric.
different size.
The input queue length versus offered load has
been depicted in Fig. 4. One can observe that the
queue length at the input of switching fabric
increases nonlinearly along with that increase of
traffic load and size of a fabric (MIN), but that
increase intensifies for greater values of both
parameters. This is result of packet collisions and
lower probability to cross over a fabric (i.e. to be
switched at the fabric output).
Usefulness of switching fabric in various
scenario of its operation mode is illustrated in Fig. 5,
i.e. the operation of switching fabric has been
examined in both a circuit and a packet mode. In
both cases diverse control rules were used for
SEs of the fabric, i.e. distributed and two versions
versions (shift and flip) of centralized control as
well. Shift control means a group of SEs governed
by a single control signal as opposite to the flip
control whereby a single signal controls the whole
stage of the cube type MIN.
It is clear (cf. Fig. 5) that usefulness of
centralized control is questionable, because in both
operation modes of the fabric such a control
mechanism decreases fabric performance (i.e. for the
PerformanceofSwitchingFabricsUsedfortheScalableRouter
569
0 . 0 0 . 2 0 . 4 0 . 6 0.8 1 . 0
coefficient of fabric utilization
N = 8
circuit mode b
we
= 5
packet mode
(buffered switching fabric)
distribut.
c- shift
c- flip
distribut.
c-shift
c-flip
0 . 0
0 . 1
0 . 2
0 . 3
0 . 4
0 . 5
0 . 6
g
x
offered
load
Figure 5: Usefulness of switching fabric in various
scenario of its operation mode.
shift control, ~41% reduction in a circuit mode and
~25% for a packet mode; even worse performance is
noticed for the flip mechanism). Hence, one can
claim that the usage of MIN with centralized control
can result in the congestion in switching fabric, i.e.
such control mechanism can have a detrimental
effect on the performance measures.
0 . 0 0 . 2 0 . 4 0 . 6 0 . 8 1 . 0
0
1 0
2 0
3 0
4 0
5 0
traffic loss
*
N
=
8
crossbar
packet mode
(buffered switching fabric)
circuit mode b
we
= 5
PM2i - distr
g
x
crossbar
selector
selector
offered
load
Figure 6: Packet loss vs. traffic load in the PM2I type MIN
for various modes of SEs’ operation.
On the other hand, there are architectures (the PM2I
type MINs) which already have built-in redundancy
(multiple paths between a given network input-
output pair) that provides fault tolerance. Therefore,
it was interesting to know what would be the effects
on fabric performance when using different types of
SEs (that allow for diverse switching functions). In
consecutive tests, the impact of different SE type has
been evaluated (cf. Fig. 6). One can observe
a qualitatively similar nature of the performance
measures (e.g. traffic loss) for both operation modes
of the fabric. The traffic loss is higher if SEs can
perform only one (selector) switching function, and
the loss is lesser if all switching functions can be
performed by SE simultaneously. This phenomenon
is more clear when traffic load increases and it is
result of the cumulative effect of events occurring
during a packet cross over the switching fabric.
Furthermore, it was also interesting to know what is
the efficiency of switching when a fabric is based on
various MINs. This aspect of performance
evaluation has been depicted in Fig. 7.
Figure 7: Efficiency of switching vs. traffic load for
a fabric based on various MINs.
It is clear that capabilities of SEs as well as the
architecture of switching network (MIN) can have
a crucial impact on the fabric performance.
According to expectations, the efficiency of
switching deteriorates along with that increase of
traffic load offered to the fabric, however the
decrease depends on the switching capabilities and
features of MINs.
5 CONCLUSIONS
This paper presents the results of performance
evaluation of the single panel multistage switching
fabrics. Two classes of packet switched MINs, those
with redundant interconnection paths (the PM2I type
MINs) and those without (the cube type MINs), have
been examined by simulated experiments in diverse
traffic conditions. Simulations under different
scenarios (including the SE model and operation
mode of switching fabric) have shown that the
performance of multistage switching fabrics
deteriorates after an increase of load. However, the
switch based on the PM2I type MIN outperforms
that based on the cube type MIN under any traffic
scenario. Moreover, presented simulation results
should be also helpful in predicting the performance
vs. SE capability trade-off before actual fabrication
of the switching fabric. The simulation also revealed
that fault tolerance of the PM2I type multistage
switching fabrics can be very useful for scalable
routers, i.e. such single panel multistage switching
SIMULTECH2014-4thInternationalConferenceonSimulationandModelingMethodologies,Technologiesand
Applications
570
fabrics seem to be an attractive alternative to the
multiple panel architecture of switching networks.
ACKNOWLEDGEMENTS
This work has been funded by the AGH University
of Science and Technology, Kraków, within the
statutory research grant.
REFERENCES
Chao, H. J., Liu, B., 2007. High Performance Switches
and Routers. J. Wiley.
Hulicki, Z., 2013. ‘Performance aspects of some switching
networks’. Proc. SIMULTECH 2013, Jul. 2013,
Reykjavik, Iceland, pp. 90-93.
Hulicki, Z., 2013a. ‘Performance analysis of multistage
switching networks’. Proc. IEEE ATNAC 2013, Nov.
2013, Christchurch, New Zealand, pp. 52-56.
Lien, Ch.-M., Chang, Ch.-S., Cheng, J., Lee D.-S., Liao,
J.-T., 2010. ‘Using banyan networks for load-balanced
switches with incremental update’, GLOBECOM
Workshops, 2010 IEEE, pp. 1–6.
Rongsen, H., Delgado-Frias, J.G., 2007. ‘Fault tolerant
interleaved switching fabrics for scalable high-
performance routers’, IEEE Trans. Paral. Distrib.
Syst., vol. 18, no. 12, Dec. 2007, pp. 1727-1738.
Skalis, N., Mhamd, L., 2010. ‘Performance guarantees in
partially buffered crossbar switches’, GLOBECOM
Workshops, 2010 IEEE, pp. 1–6.
Tzeng, N.-F., 2004. ‘Multistage-based switching fabrics
for scalable routers’, IEEE Trans. Parallel and
Distributed Systems, vol. 15, no. 4, pp. 304-318.
PerformanceofSwitchingFabricsUsedfortheScalableRouter
571
