Optimizing storage performance in public cloud platforms

Cloud computing is an elastic computing model where users can lease computing and storage resources on demand from a remote infrastructure. It is gaining popularity due to its low cost, high reliability, and wide availability. With the emergence of public cloud storage platforms like Amazon, Microsoft, and Google, individual applications and enterprise storage are being deployed on Clouds. However, a serious impediment to its wider deployment is the relative lack of effective data management services. Our experiments, as well as industry reports, have shown that the performance and service-level agreement (SLA) cannot be guaranteed when the data is served over public Clouds. The relatively slow access to persistent data and large variability in cloud storage I/O performance can significantly degrade the performance of data-intensive applications. This paper addresses the issue of I/O performance fluctuation over public cloud platforms and we propose a middleware called CloudMW between the Cloud storage and clients to provide the storage services with better performance and SLA satisfaction. Some technologies, including data virtualization, data chunking, caching, and replication, are integrated into CloudMW to achieve a more stable and predictable performance, and permit flexible sharing of storage among the virtual machines (VMs). Experimental results based on Amazon Web Services (AWS) show that CloudMW is able to improve the stability and help provide better SLAs and data sharing for cloud storage.

Sourcing Decisions with Capacity Reservations under Supply Disruptions

Firms employing both offshore outsourcing and nearshore sourcing strategies may face supply disruption,demand uncertainty,and quality risks simultaneously.Sourcing decisions become inevitably important and complicated when both profit and the customer-service level are taken into consideration.In this paper,we model a scenario where a manufacturer who faces stochastic demand procures major modules from an overseas supplier and two local suppliers.The overseas supplier offers quality products while being susceptible to disruption risks;if the local suppliers,who are completely reliable and serve as a backup,offer products that are of inferior quality,it may result in lower market acceptance and a bad experience for the final customers.The manufacturer has to reserve capacity with backup suppliers before urgent orders are placed,when the primary source experiences a shortfall.We explicitly derive the manufacturer’s optimal order quantities and reservation quantities,which are functions of the heterogeneous suppliers’wholesale prices,reservation prices,and other parameters.The impacts of the fill-rate constraint and customer-experience quality constraint on the manufacturer’s purchasing decisions are investigated.Interesting managerial insights on the merits of backup sourcing with capacity reservations for managing demand uncertainties and supply disruption risks are also discussed.

Automated Power Control for Virtualized Infrastructures

Power control for virtualized enviromnents has is keeping underlying infrastructure in reasonably low power gained much attention recently. One of the major challenges states and achieving service-level objectives （SLOs） of upper applications as well. Existing solutions, however, cannot effectively tackle this problem for virtualized environments. In this paper, we propose an automated power control solution for such scenarios in hope of making some progress. The major advantage of our solution is being able to precisely control the CPU frequency levels of a physical environment and the CPU power allocations among virtual machines with respect to the SLOs of multiple applications. Based on control theory and online model estimation, our solution can adapt to the variations of application power demands. Additionally, our solution can simultaneously manage the CPU power control for all virtual machines according to their dependencies at either the application-level or the infrastructure-level. The experimental evaluation demonstrates that our solution outperforms three state-of-the-art methods in terms of achieving the application SLOs with low infrastructure power consumption.