Workshop Program

Cloud computing realises the vision of utility computing. Tenants can benefit from on-demand provisioning of computational resources according to a pay-per-use model and can outsource hardware purchases and maintenance. Tenants, however, have only limited visibility and control over network resources. Even for simple tasks, tenants must resort to inefficient overlay networks.

To address these shortcomings, we propose Network-as-a-Service (NaaS), a framework that integrates current cloud computing offerings with direct, yet secure, tenant access to the network infrastructure. Using NaaS, tenants can easily deploy custom routing and multicast protocols. Further, by modifying the content of packets on-path, they can efficiently implement advanced network services, such as in-network data aggregation, redundancy elimination and smart caching.

We discuss applications that can benefit from NaaS, motivate the functionality required by NaaS, and sketch a possible implementation and programming model that can be supported by current technology. Our initial simulation study suggests that, even with limited processing capability at network switches, NaaS can significantly increase application throughput and reduce network traffic.

We introduce the notion of a Personal Cloud — a collection of Virtual Machines (VMs) running on unused computers at the edge. The Personal Cloud provides an ideal solution for the secure sharing of compute and storage resources across peers in a resource and application agnostic manner, and facilitates new computational paradigms such as datacenter-less, distributed virtual clouds. We provide and implement solutions for the challenges of managing a Personal Cloud, such as IP address sharing, bandwidth sharing and isolation from local home network traffic. We also propose and implement a provably optimal solution to the resource management problem, allowing peers to share VMs across their individual Personal Clouds by specifying their resource offers and requests, and verify its performance via detailed simulations.

Offloading has emerged as a promising idea to allow handheld devices to access intensive applications without performance or energy costs. This could be particularly useful for enterprises seeking to run line-of-business applications on handhelds. However, we must address two practical roadblocks in order to make offloading amenable for enterprises: (i) ensuring data privacy and the use of trusted offloading resources, and (ii) accommodating offload at scale with diverse handheld objectives and compute resource capabilities. We present the design and implementation of an Enterprise-Centric Offloading System (ECOS) which augments prior offloading proposals to address these issues. ECOS uses a logically central controller to opportunistically leverage diverse compute resources, while tightly controlling where specific applications offload depending on privacy, performance, and energy constraints of users and applications. A wide range of experiments using a real prototype establish the effectiveness of our approach.

Automatically managing collocated cloud applications for improved performance is a hard problem due to the unprecedented scale and the dynamics of the multiplexed cloud environment. Compounding the problem, today’s approaches to cloud application management are too limited in the way they acquire information. Monitoring performed by the operator is too low level and application agnostic while monitoring performed by applications in isolation is too restricted. In this paper 1 we propose sharing of application experiences as a cloud and application management building block. To achieve this level of sharing, the current state-of-the-art towards increased isolation among cloud applications needs to be re-thought. Focusing on isolation overlooks and potentially impedes the substantial benefits obtainable through sharing application experiences. We explore the benefits, challenges and incentives associated with sharing application experiences and argue why sharing is a winning proposition for both operators and applications. We also propose a web portal synergy2cloud.com that we hope will serve as a stepping stone for making cross-sharing cloud application experiences a reality.

As cloud service providers leverage server virtualization to host applications in virtual machines (VMs), they must ensure proper allocation of resource capacities in order to satisfy the contracted service level agreements (SLAs) with the application owners. However, the ever-growing number of virtual and physical machines within such infrastructure creates greater challenges in quickly and effectively localizing the system bottlenecks that lead to SLA violations. This paper describes DAPA, a new performance diagnostic framework to help system administrators analyze application performance anomalies and identify potential causes of SLA violations. DAPA incorporates several customized statistical techniques to capture the quantitative relationship between the application performance and virtualized system metrics. We have built a prototype implementation of DAPA on a cluster of virtualized systems to diagnose a set of SLA violations for an enterprise application. Preliminary evaluation results show that DAPA is able to localize the most suspicious attributes of the virtual machines and physical hosts that are related to the SLA violations.

Cloud computing leverages virtualization to offer resources on demand to multiple “tenants”. However, sharing the server and network infrastructure creates new vulnerabilities, where one tenant can attack another by compromising the underlying hypervisor. We design a system that supports virtualized networking using software switches without a hypervisor. In our architecture, the software switch runs in a Switch Domain (DomS) that is separate from the control VM. Both the guest VMs and DomS run directly on the server hardware, with processing and memory resources allocated in advance. Each guest VM interacts with the software switch through a shared memory region using periodic polling to detect network packets. The communication does not involve the hypervisor or the control VM. In addition, any software bugs that crash the software switch do not crash the rest of the system, and a crashed switch can be easily rebooted. Experiments with our initial prototype, built using Xen and Open vSwitch, show that the combination of shared pages and polling offers reasonable performance compared to conventional hypervisor-based solutions.