Performance Evaluation and Improvement of Next-generation Optical Switching Architecture
Abstract
The advance in wavelength-division multiplexing (WDM) technology, which
propagates closer to the network edge, enables the proliferation of applications with
very high bandwidth and stringent performance requirements. However, today’s core
networks, which mainly employ electronics for bulk processing and switching,
eventually will become strained in both capacity and flexibility. The introduction of
optical technologies in switching functions is a major evolution of core network. To
relieve electronics switching, optical concepts like wavelength, waveband, and even
fiber switching were developed and are currently introduced, e.g. by reconfigurable
optical add-drop multiplexers (i.e. ROADM) and all-optical cross-connects (i.e. OXC).
Optical switching methods have many distinct advantages compared with electrical
switches: removing the expensive Optical-Electronic-Optical (OEO) conversion;
providing high switching speed; benefiting environment with reduced power
consumption; being independent with protocol and bit rates of the network. However,
the clear and concrete solution of the switching architecture for the next generation
optical networks has not been decided. In this thesis, a switching architecture named 3-
Level-Integrated-Hybrid-Optical-Network (i.e. 3LIHON) is introduced as a potential
candidate for the next-generation optical networks. Its virtue is related to the
combination of merits of three different switching paradigms, which have different
switching granularities. Accordingly three different traffic transport types (i.e., service
classes) are transmitted on the same wavelength in a time-interleaved way. Thus,
3LIHON network can utilize the network resources efficiently while supporting a wide
range of service classes with different service demands.
3LIHON concept is first proposed in proceedings of international conference IEEE
GLOBECOM in 2011. Its novel concept opens wide variable research scenarios,
ranging from the architecture realization and implementation (e.g. the detailed switching
subsystem architectures definition and the respective working principles), to
performance and availability evaluation (e.g. the performance metrics of each service
class under the employed working principles and the architecture), as well as overall
network architecture cost and power consumption analysis. The 3LIHON concept is the
starting point of this PhD thesis. This thesis focuses on three research directions based
on 3LIHON concept: first is the 3LIHON subsystem architecture design; second is the
performance evaluation of the considered architecture; third is the performance
modelling of employed working principles inside 3LIHON network (e.g. the preemptive
scheduling which is widely used inside 3LIHON network-queue scheduling
and part of output contention resolution mechanism). The ten included papers respond
to these three directions and form the main contributions of this thesis. Among these
contributions three concrete topics are treated in this thesis, i.e. the architecture design of 3LIHON edge node, OPS subsystem definition and improvement, the scheduling
algorithm inside 3LIHON network and the respective performance modelling.
The 3LIHON concept defines three service classes which are served by different
switch subsystems: Guaranteed Service Type (i.e. GST); Statistically
Multiplexing/Real-Time service type (i.e. SM/RT); Statistically Multiplexing/Best-
Effort service type (i.e. SM/BE). GST service, which covers high real-time and
bandwidth demands applications, is handled by OXC. OXC is the best choice for large
scale high performance demands applications, due to its resource reservation and
monopoly. By being handled by OXC, the strict performance of GST service can be
guaranteed. SM/RT service, corresponding to very low bandwidth but very high real
time demands applications, is switched by OPS equipped with limited or no buffer, due
to its high switching speed and extremely short delay time. SM/BE service, which
covers low real time demands applications, is served by EPS (i.e. electronic packet
switching) which employs electronic buffers to reduce the respective packet loss as
much as possible. However, how to map the incoming traffic into the three predefined
service classes, which is operated by 3LIHON edge node, is not given by 3LIHON
concept. Thus the first research topic of this PhD thesis is to design the 3LIHON edge
node. The respective architecture (e.g. how to add up the traffic into 3LIHON network)
and detailed working principles (e.g. the assembly algorithm of the GST traffic, the
queueing scheduling for each buffer subsystem and the output contention resolution
mechanism) are given in this thesis. The proposed architecture employs multiple
electronic buffers to add up the local traffic into 3LIHON network whereas whether the
cross traffic go through the electronic buffers depends on the service type. For GST,
which is transmitted in circuit switched mode, no contention exists between local and
transit GST traffic. Thus only the local GST traffic goes through the electronic buffers.
We introduce the MBMAP (e.g. Max-Burstlength-Max-Assembly-Period) assembly
algorithm to produce the GST burst. For SM/RT, which is very sensitive to packet delay
and packet loss, we give the high priority to transit SM/RT traffic which is not buffered.
The local SM/RT traffic goes through very limited electronic buffers in the 3LIHON
edge node to reduce the packet loss introduced in the switch input. For SM/BE, which
has high tolerance toward delay time, the sufficient size electronic buffers are employed
in edge node to reduce its packet loss. For the switch output, where three service classes
compete for the same wavelengths, the hierarchical contention resolution mechanism is
applied. The GST service has the non-preemptive priority over SM/RT service, whereas
both GST and SM/RT have preemptive priority over SM/BE service. For the
performance evaluation of this proposed edge node architecture, six source types and
two traffic types are applied to model the traffic inside the network as realistic as
possible. This thesis mainly applies Demos/Simula to execute the simulations. The
results showed that the proposed 3LIHON edge node architecture provides desirable performance for each service class. The detailed architecture and parameters’ design
guidelines are given according to the results.
The second topic in this thesis is the OPS subsystem definition. For 3LIHON
network, the OPS subsystem is mainly used for switching very high real-time and low
bandwidth demand SM/RT traffic. Thus the potential OPS candidate should have
no/limited buffers to reduce the introduced delay. And the economic cost of OPS (e.g.
the small physic size and the mature optical components) should be made as cheap as
possible due to its extremely low requirement of throughput. Till now a lot of OPS
architectures have been proposed. However, most of them focus on the packet
contention resolution, fast packet switching and QoS. Not much attention is paid
regarding the maturity of optical components and the corresponding economic cost,
which are also major issues for the OPS subsystem of 3LIHON network. In order to
achieve the best OPS candidate, we first evaluate the existing OPS architectures.
Among them, the well-known OPS architecture equipped with limited range wavelength
converters (i.e. OPS-LRWC) attracts research efforts. Compared with the general OPS
with full wavelength converters (i.e. OPS-FWC), it is more promising due to its smaller
physical size, less economic cost and easier implementation. To evaluate its
performance, two separate Markov models are built to evaluate the packet loss of OPSLRWC
under different network scenarios. The upper and lower packet loss values are
derived directly by these two models. The results show that these two models provide
desirable packet loss bounds for OPS-LRWC, independent of the employed conversion
policies. Furthermore, the applicability of the proposed models depends on the
normalized system load, the conversion range of the LRWCs, and the employed
conversion policy. In addition, the results show that the packet loss decreases as the
conversion range increases. However, different from the synchronous OPS networks,
which have a threshold of the conversion range for limiting the performance
improvement, no choice of the best wavelength conversion range exists in the
asynchronous OPS networks. In order to find the best candidate, we also propose a
novel OPS architecture equipped with input concentrators before the switch fabric.
Multiple cascaded fiber delay lines are used to build the input concentrators, the main
function of which is to aggregate the incoming traffic from same wavelength into only
one input link of switch fabric. Accordingly, the physical size and the respective
economic cost of OPS are reduced significantly, whereas the link utilization increases
dramatically. This architecture is designed special for very low incoming traffic
scenario, e.g. 3LIHON network scenario. However, the employment of input
concentrators would introduce the additional delay time in the switch input. In order to
evaluate its performance, we build a one-dimensional Markov model according to the
packet departure process from the input concentrator. The packet loss and mean packet
delay introduced by input concentrators can be given by Markov model directly. The results show that the additional delay time is far less than the tolerant delay time.
Meanwhile a learning curve model is introduced to analyze its economic cost, which is
expressed as a function of time. The proposed OPS architecture is firstly compared with
OPS without input concentrators based on AWG, and then compared with the OPSLRWC
architectures based on B&S switch fabric (i.e. broadcast and select). Both
results show that the proposed OPS architecture is much cheaper while providing same
performance as these architectures.
Some more works has been done to improve the proposed OPS architecture, in which
the input concentrators dominate the performance of whole OPS node. We generalize
the input concentrators as multiple input single output delay-line optical buffers. A
novel modeling approach is developed to denote the delay time evolution of each new
arrival to the buffer. Both fixed (all packets’ length is constant) and variable length (the
packet length follows n.e.d. distribution) incoming traffic are considered. This approach
is validated by simulations under different network scenarios. The results show that the
basic FDL length is a very important parameter for the buffer performance. The optimal
FDL length value is dominated by incoming traffic load and the packet length (mean
packet length for case-n.e.d. incoming packets), and independent of the buffer size
offered by buffer. These results validate that setting the FDL length same as the packet
length is not the best choice for delay-line buffers. Meanwhile, this approach gives the
guidelines for designing the optical buffer with the best performance under different
scenarios. By adjusting the basic FDL length value, the performance of the whole OPS
node can be improved significantly. It is noticeable that all results achieved from this
topic can not only be used in 3LIHON network, but also very useful for general optical
network scenarios, e.g. the proposed OPS architecture can widely be employed in the
access networks with low entrance traffic; the evaluation results from the proposed
input concentrators can generally be used for improving the general optical networks
with such delay-line buffers.
The third topic of this thesis is the evaluation of preemptive scheduling, which is
widely used in optical networks as service differentiation. In 3LIHON network, the
hierarchical contention resolution mechanism is employed at the switch output, in
which both of GST and SM/RT classes have the preemptive priority over SM/BE.
Meanwhile, the preemptive scheduling is generally used as contention resolution in pure
OPS/OBS networks. Thus in this thesis we pay special attention to study the preemptive
scheduling. Its performance and operations are evaluated in both optical networks and
the general multi-service systems. Three included papers (i.e. P6, P7 and P8) contribute
to this topic. For performance modeling, a novel approximation model-hierarchical
Markov model is proposed. According to the priority of each class, the model is built as
multiple levels. Each level presents all the transmission states of the respective service
class. When calculating the blocking probabilities, conditional probability principles are used to weigh and sum contributions from each one-dimensional Markov chain in each
level. This model only needs to increase the number of levels when modeling the larger
system with more classes, thus it shows excellent extensibility. Compared with the
general analytical modeling, which use the multi-dimensional Markov model to analyze
the small system with less classes and resources, the proposed model has several major
advantages: applicability of analyzing the general multi-service systems, closed-form
expressions of the blocking probabilities and the significantly decreased computational
complexity. Furthermore, this thesis validates the accuracy of the proposed hierarchical
model. For the service class with highest priority, this model always provides the
accurate packet loss. However for other services, this model provides the satisfactory
approximate results. The respective discrepancy decreases as the relative load allocation
of all higher service classes (compared with the studied service class) decreases. In
addition, it is noticeable that, all results obtained from the proposed hierarchical Markov
model are applicable for the general multi-service systems with preemptive scheduling