Monitoring and Adaptation of Pooled Cloud Computing Resources

Cloud computing service providers seek to deploy vast quantities of shared resources that
are managed in an optimal way for all consumers. Monitoring describes a collection of
techniques for measuring cloud resource consumption and performance, and for detecting
anomalies. Adaptation describes a collection of techniques for effecting changes to a cloud
configuration in order to improve performance, or counter anomalous behaviour.

The state-of-the-art in cloud monitoring mostly generates statistics about Infrastructureas-
a-Service (IaaS) which is used privately to inform the providers about the health of
the service. Consumers see other aspects, such as service latency and response-times,
affected by numbers of tenants on a given platform. While cloud providers offer Service
Level Agreements (SLAs), it is often not possible to tell whether these are being honoured.
Providers want to predict performance, plan capacity and detect anomalies well in advance;
and consumers need guarantees that their SLAs are being honoured.

Our approach presented a real cloud testbed with the capabilities of proactively monitoring
and gathering cloud resources information for making prediction and forecasting. As part
of this, new adaptive filtering approaches were explored at the heart of the feedback driven
control algorithms. This research investigated the use of several adaptive filtering techniques,
including the Least Mean Square (LMS) as a benchmark, but later focused on the Kalman
optimal H2 filter, and the boosted decision tree regression machine learning method, as
possibly useful algorithms for tracking resource consumption and making predictions about
future resource usage. Our results show that two regression methods can model the problem
of cloud resource prediction based on past observations. The Boosted decision tree produced
the most suitable predictions based on the feature labels such as the number of virtual users.
Our investigations also reveal that the optimal Kalman filter can predict and forecast cloud
resource consumptions into a future time horizon accurately. We also show that this predictive
analytic framework is inline with the concept of the MAPE-K in autonomous computing.