MORAIS, F. J. A.; http://lattes.cnpq.br/0987042606840444; MORAIS, Fábio Jorge Almeida.
Abstract:
The paradigm of cloud computing has brought the concept of elasticity, which is the on
demand provision of computational resources. This property is the key to reduce the costs
derived from the execution of services in cloud systems that employ an infrastructure-as-aservice
(IaaS) deployment model. However, this property can only be fully exploited if the
users of IaaS services are able to estimate the short-term future demands of their own applications,
so that only the necessary infrastructure to maintain the application is requested at
each instant of time. Thus, the service level agreements (SLA), signed between the client of
the IaaS service and the users of their applications, are always honored and over provisioning
is avoided.
The auto-scaling is the process of dynamically modifying the amount of resources available
to maintain and run an application on an IaaS system, depending on the load application.
The state-of-practice provides solutions for auto-scaling using reactive approaches, which
are not sufficient to minimize the costs of SLA violations, although they may reduce the
costs of over provisioning. To reduce costs due to SLA violations proactive approaches are
necessary.
This work proposes a flexible non-intrusive framework for auto-scaling services. The
framework follows a hybrid, reactive and proactive, approache based on the use of a configurable
set of predictors for the future demand of services and uses a selection mechanism
that decides, over time, the best predictor to be used. Also a new way of correcting underestimations
is proposed, which reduces the number of SLA violations.
The proposed framework had its performance evaluated through simulations using production
utilization traces of HP customers. The results show that costs savings of as much
as 37% can be achieved, while the probability of an SLA violation can be kept, on average,
as small as 0.008%, and no larger than 0.036%. Moreover, the flexibility of the framework
allows different configurations to be used, for which additional cost savings can be achieved
with only a small increase on the number of SLA violations.