Efficient Algorithms for Dynamic Cloud Resource Provisioning

Abstract

Cloud computing has emerged as a new computing paradigm, with data centers proliferating in today’s Internet. Cloud service providers often adopt static resource provisioning to pack cloud resources to fixed types of virtual machines (VM), failing to address user demands efficiently and precisely. In this thesis, we focus on dynamic cloud resource provisioning, which provides realtime, on-demand access to cloud resources. We propose efficient algorithms to guide resource allocation and workload dispatching in cloud systems. We first study dynamic VM provisioning via an online auction algorithm. We generalize the existing literature by introducing computing jobs with completion deadlines. A cloud user bids for future cloud resources to execute its job. Each bid specifies (a) a resource profile of tailor-made VMs, (b) a utility, reflecting the amount that the user is willing to pay for executing its job, and (c) a soft deadline, specifying the preferred finish time of the job, as well as a penalty function that characterizes the cost of violating the deadline. We propose efficient cloud job auctions that execute in an online fashion, provide truthfulness guarantee, and achieve a good competitive ratio. We then discuss cloud container services, a more recent form of cloud resource provisioning. Compared to traditional VMs, cloud containers are more flexible and lightweight. We exploit this new algorithm design space, and study dynamic cloud container provisioning. We design efficient scheduling algorithms for complex computing jobs that are running on cloud containers. Our offline and online schedulers permit partial execution, allow a job to specify its job deadline, desired cloud containers, and inter-container dependence relations, and achieve near-optimal expected objective values. We further extend our study to cloud container clusters. Enterprise users often create clusters of inter-connected containers to provision complex services. Compared to traditional cloud services, key challenges in container cluster (CC) provisioning lie in the optimal placement of containers while considering inter-container traffic in a CC. The challenge further escalates when CCs are provisioned in an online fashion upon CC request arrivals. We investigate dynamic cloud CC provisioning, and propose an online algorithm to address the above challenges. Our online algorithm achieves computational and economical efficiencies.