A methodological framework for cloud resource provisioning and scheduling of data parallel applications under uncertainty

Data parallel applications are being extensively deployed in cloud environments
because of the possibility of dynamically provisioning storage and computation resources. To identify cost-effective solutions that satisfy the desired service levels,
resource provisioning and scheduling play a critical role. Nevertheless, the unpredictable behavior of cloud performance makes the estimation of the resources actually needed quite complex. In this paper we propose a provisioning and scheduling
framework that explicitly tackles uncertainties and performance variability of the
cloud infrastructure and of the workload. This framework allows cloud users to estimate in advance, i.e., prior to the actual execution of the applications, the resource
settings that cope with uncertainty. We formulate an optimization problem where
the characteristics not perfectly known or affected by uncertain phenomena are
represented as random variables modeled by the corresponding probability distributions. Provisioning and scheduling decisions – while optimizing various metrics,
such as monetary leasing costs of cloud resources and application execution time –
take fully account of uncertainties encountered in cloud environments. To test our framework, we consider data parallel applications characterized by a deadline constraint and we investigate the impact of their characteristics and of the variability
of the cloud infrastructure. The experiments show that the resource provisioning
and scheduling plans identified by our approach nicely cope with uncertainties and
ensure that the application deadline is satisfied.