Dynamic Resource Management for Cloud-hosted Internet Applications

Download or read book Dynamic Resource Management for Cloud-hosted Internet Applications written by Hangwei Qian and published by LAP Lambert Academic Publishing. This book was released on 2012 with total page 132 pages. Available in PDF, EPUB and Kindle. Book excerpt: Internet is evolving toward service-oriented computing platforms (e.g., cloud computing platforms, such as Amazon EC2 and Microsoft Azure). In these platforms, service providers (owners of the platforms) offer resource pools by building multiple geo-distributed data centers; application providers (owners of the applications) outsource the hosting of their applications to these platforms, and pay by the amount of resources used as utility. These multi-tenant platforms need to dynamically allocate resources to applications so as to meet their demand variation. In this thesis, we address several issues of the dynamic resource management in these platforms. On the one hand, we consider the resource provisioning problems within data centers. In order to allocate resources to applications quickly, we propose deploying ghost virtual machines (VMs) which host spare application instances across the physical machines. When an application needs more instances, we can configure the request distributer to forward requests to ghost VMs, which takes only 5-7 seconds. Also, to deal with the scalability issues in mega data center (with hundreds of thousands of servers), we introduce hierarchical resource management scheme in which servers are divided into groups (pods), each with about 5k servers, and existing techniques are employed to manage resources in each pod efficiently. Meanwhile, multiple strategies are explored to balance the load among the pods. In addition, we also propose a new data center architecture in which we can apply DNS-based mechanism to balance the load among the access links which connect data center to Internet. On the other hand, we address the resource management problems among multiple data centers. We proposed a unified approach to decide in how many/which data centers each application should be deployed, and how client requests are forwarded to the geo-distributed service replicas. We make these decisions based on a min-cost network flow model, and apply a novel demand clustering technique to overcome the scalability issue when solving the min-cost problem. Furthermore, we also introduce a new client-side DNS architecture which brings local DNS server close to clients so that DNS-based server selection can precisely choose close service replicas for clients.