Cloud-service agreements entail the provisioning of virtual infrastructures at specified availability levels. The agreements specify the infrastructure prices and penalties borne by providers when assured availabilities are not met. Hence, backup resources should be optimally provisioned by balancing the provisioning costs with penalties. We develop stochastic dynamic optimization models of the backup provisioning problem by integrating methods from Operations Management, Statistics and Machine Learning, and Computer Science. We present two sets of dynamic provisioning strategies: periodic policies where resources are adjusted at regular intervals and aperiodic policies that allow flexible timing of such interventions. A closed-loop optimization model under conservative resource control and a certainty-equivalent optimization model under aggressive resource control are developed for periodic management. Similarly, we propose two strategies for aperiodic management: single intervention with single look-ahead and multiple interventions with single look-ahead. We develop computationally efficient online algorithms for both periodic and aperiodic models. We evaluate the models through computational experiments and validate with Amazon EC2 use cases. We provide cloud service providers practical guidelines for efficiently formulating flexible service agreements and cost-effectively managing the virtual infrastructure resources. The proposed resource optimization framework is most suitable to IaaS and can easily be extended to PaaS environments.

Energy costs represent a significant share of the total cost of ownership in high-performance computing systems. Using a unique data set collected by massive sensor networks in a petascale national supercomputing center, we first present an explanatory model to identify key factors affecting energy consumption in supercomputing. Our analytic results show that workload distribution among the nodes has significant effects and could effectively be leveraged to improve energy efficiency. We then establish the high model performance using in-sample and out-of-sample analyses and develop prescriptive models for energy-optimal runtime workload management. We present four dynamic resource management methodologies (packing, load balancing, threshold-based switching, and energy optimization), model their application at two levels (within-rack and cross-rack resource allocation), and explore runtime resource redistribution policies for jobs under the computational steering and comparatively evaluate strategies that use computational steering with those that do not. Our experimental results lead to a threshold strategy that yields near-optimal energy efficiency under all workload conditions. We further calibrate the energy-optimal resource allocations over the full range of workloads and present a bi-criteria evaluation to consider energy consumption and job performance tradeoffs. We conclude with implementation guidelines and policy insights into energy-efficient computing resource management in large supercomputing centers.

Cloud systems, which are typical cyber-physical systems, consist of physical nodes and virtualized facilities that collaborate to fulfill cloud computing services. The advent of visualization technology engenders resource sharing and service parallelism in cloud services, introducing novel challenges to system modeling. In this study, we construct a systematic model that concurrently evaluates system reliability, performance, and power consumption (PC) while delineating cloud service disruptions arising from random hardware and software failures. Initially, we depict system states using a birth-death process that accommodates resource sharing and service parallelism. Given the relatively concise service duration and regular failure distributions, we employ transient-state transition probabilities instead of steady-state analysis. The birth-death process effectively links system reliability, performance, and PC through service durations governed by service assignment decisions and failure/repair distributions. Subsequently, we devise a multistage sample path randomization method to estimate system metrics and other factors related to service availability. The findings highlight that the trade-off between performance and PC, under the umbrella of reliability guarantees, hinges on the equilibrium between service duration and unit power. To further delve into the subject, we formulate optimization models for service assignment and juxtapose optimal decisions under varying availability scenarios, workload levels, and service attributes. Numerical results indicate that service parallelism can improve performance and conserve energy when the workload remains moderate. However, as the workload escalates, the repercussions of resource sharing-induced performance loss become more pronounced due to resource capacity limitations. In cases where system availability is constrained, resource sharing should be approached cautiously to ensure adherence to deadline requirements. This study theoretically analyzes the interrelations among system reliability, performance, and PC, offering valuable insights for making informed decisions in cloud service assignments.

Network function virtualization enables efficient cloud-resource planning by virtualizing network services and applications into software running on commodity servers. A cloud-service provider need...