Modelling and Analyzing Cost-Effective Dependability in Passive Optical Networks
Abstract
Lately, the proliferation of new applications and services as well as the
penetration of the Internet has created a growing need for the deployment
of broadband access networks. Mostly due to the high bandwidth they
offer, fiber-based solutions are well regarded to implement both present
and future broadband access networks. Hence, not only operators and
providers are upgrading their access networks to fiber, but also regulators
and governmental institutions are supporting its deployment to meet the
digital demands of our society. Particularly, Passive Optical Networks
(PONs) are one of the most important types of fixed fiber-based access
systems today, being widely deployed worldwide. Besides, PONs are
also key to enable Fiber-Wireless (FiWi) integration, serving as backhaul
networks. Being fiber-based, PONs offer high bandwidth, but also a set
of features, such as low-power consumption or easy management, that
make them very attractive as broadband access networks.
Nevertheless, real telecommunication networks are not failure free.
With society, industry and individuals becoming more dependent on
telecommunications networks, users also demand these networks to provide
dependable access to applications and services. In a dynamic business
ecosystem, being able to meet user demands is key to achieve a successful
network business. Further, in this context, dependability plays a major
role in the overall economics of a business. First, achieving dependability
entails a cost, typically due to the investment in fault tolerance. This
is especially a concern in access networks, which are more cost sensitive
than other parts of the network, due to the lower number of served users.
On the other hand, poor dependability has consequences that negatively
impact a business, from direct cost of failures and repairs, to bad publicity
and loss of reputation. It is then a challenge for operators to increase the
dependability of their access networks, hence reducing the consequences
of dependability, at an affordable cost. Consequently, this thesis aims at
addressing the trade-off between the cost of achieving dependability and
its consequences, i.e. cost-effective dependability, in PON deployments.
As a first step, this work focuses on modelling the PON deployment
area, which affects both sides of the trade-off, due to infrastructure
sharing, client clustering and design decisions. Especially, this suggestion
builds on a network geometric model, i.e. the Manhattan model, to
develop a closed formulation for the Capital Expenditures (CAPEX),
accounting for the cost of achieving dependability. Additionally, this
formulation also allows for calculating the failure impact (the number
of clients affected by a failure) and thus capturing failure dependencies among clients. With the physical framework provided by the Manhattan
model and the related equations, the CAPEX, asymptotic availability and
failure impact of both unprotected and protected PONs can be analyzed.
A central part of the thesis focuses on how to model the consequences
of dependability, covering penalties due to breached Service Level Agreements
(SLAs), cost of maintenance, buying of spare parts for repair and
loss of reputation. Based on the constraints of the Manhattan model,
a first approach proposes the use of Markov cost models to estimate
dependability consequences fully in monetary units, i.e. as Operational
Expenditures (OPEX). By developing the necessary equations, the objective
is to express dependability attributes (asymptotic availability and
failure impact) as expected OPEX. Thus, by combining CAPEX and
OPEX, a first analysis of cost-efficient dependability with respect to
hardware failures is performed, aiming at reducing the total cost.
Also, an important part of this thesis deals with introducing software
failures into the modelling of PON dependability. To do that, the software
failure intensity of a PON Optical Line Termination (OLT) is estimated
from empirical data by employing Duane’s model for software reliability
growth. Different types of software failures are characterized, and
integrated with hardware failures in the Markovian cost model. Further,
how to model hardware-software interaction with respect to imperfect
recovery of hardware faults due to software is also presented.
Then, with the insight provided in software failures, the approach
to model the consequences of dependability is further refined. Namely,
a risk assessment approach, considering the probability distribution of
the interval availability during a finite time frame, and thus the risk
it represents, is followed. Under this approach, the full probability
distribution of the OPEX can be computed, yet loss of reputation is
modelled through client dissatisfaction and large outages. Hence, the
probability mass function of the number of dissatisfied clients, as well as
scatter plots of the down time versus the failure impact. This approach
not only gives knowledge about the stochastic behaviour of these three
aspects, but also shows how different fault tolerance mechanisms modify
this behaviour and its associated risks.
Lastly, a final consideration of this work proposes how to protect
against software failures and assign available capacity to improve the
interval availability of a PON client. In essence, a policy to assign
available capacity to clients depending on their accumulated down time is
suggested. This policy modifies the distribution of the interval availability,
improving its expectation and reducing its variability, while allowing for
the implementation of differentiated dependability.