No abstract

No abstract

Abstract-The emergence of cloud environments has made feasible the delivery of Internet-scale services by addressing a number of challenges such as live migration, fault tolerance and quality of service. However, current approaches do not tackle key issues related to cloud storage, which are of increasing importance given the enormous amount of data being produced in today's rich digital environment (e.g. by smart phones, social networks, sensors, user generated content). In this paper we present the architecture of a scalable and flexible cloud environment addressing the challenge of providing data-intensive storage cloud services through raising the abstraction level of storage, enabling data mobility across providers, allowing computational and content-centric access to storage and deploying new data-oriented mechanisms for QoS and security guarantees. We also demonstrate the added value and effectiveness of the proposed architecture through two real-life application scenarios from the healthcare and media domains.

Abstract-Cloud computing is mostly allocating resources at course grain, e.g., entire CPU cores are allocated for as long as an hour. For improving resource efficiency of clouds, the Resource-as-a-Service cloud is envisioned, which allocated resources at fraction of a core and second granularity. Despite technology enabling such infrastructure, e.g., through lightweight virtualization such as LXC or vertical elasticity in the Xen hypervisor, performance models to decide how much capacity to allocate to each application are lagging behind.In this paper, we evaluate two performance models for mean response time, a previously proposed one and a novel one. The models are tested with 3 applications in both open and closed system models. Results show that the inverse model reacts fast and remains stable for targets as low as 0.5 seconds.

The Office of Naval Research (ONR) funds research that will benefit the U. S. Navy. In managing this research, ONR must allocate resources to achieve maximum benefits within its resource constraints. This report addresses one aspect of the resource allocation problem, namely, identifying funding profiles (normalized program funds as a function of time) for programs, or groups of programs, that provide maximum adherence to some predetermined policy. Four examples are presented in this report. The first describes selection of an overall funding profile for a group of special programs called Accelerated Research Initiatives (ARIs). The profile was selected to ensure an infusion of new ARIs proportional to total budget, and the actual turnover of ARIs resulting from use of the profile is shown. The second example shows how, with the use of quadratic programming, additional constraints could be placed upon the funding profiles, if desired, and relatively stable ARI turnover would still result. The third example describes the use of quadratic programming in an experiment in which each member of a group of potential programs had maximum funding profile flexibility while obeying the group funding ceiling constraints. The fourth example describes an extension of the methodology of the third example that could be applied to determining the profiles of every program in any funding organization. OVERALL FUNDING PROFILE FOR ACCELERATED RESEARCH INITIATIVES A. Background In 1981, ONR decided to devote a fraction of its total research budget to accelerated research programs (hereafter referred to as Accelerated Research Initiatives-ARIs). The objective of these ARIs was to concentrate resources into promising areas of research for a finite period of time in order to accelerate progress in the technical fields of the ARIs. The overall ARI program was to be increased in size until it achieved a target fraction of the total research budget, and then it would remain at approximately this (steady state) target level. The ARI selection process was designed to be competitive. Internal Navy organizations would submit proposals to ONR for these ARIs, and a fraction of these proposals would be funded (see Kostoff [ 1988] for a detailed description of this competitive process). Initially, funding profiles for the individual ARIs were not specified by ONR and were at the discretion of the proposers. Consequently, the programs that won the competition for the first few years had a wide range of profiles, with the main ONR control being the first year's funds for each ARI. After the ARI programs had been in existence for about three years, projections for the annual pool of funds available for new ARIs (hereafter referred to as the margin) as a function of time showed that the margin would be highly oscillatory over time. It would range in amplitude from near-zero in some years to large amounts in other years. The desired policy was to have the annual infusion of new ARI funds proportional to the total ARI funds available, at least in the s...

Designing efficient control mechanisms to meet strict performance requirements with respect to changing workload demands without sacrificing resource efficiency remains a challenge in cloud infrastructures. A popular approach is fine-grained resource provisioning via auto-scaling mechanisms that rely on either threshold-based adaptation rules or sophisticated queuing/control-theoretic models. While it is difficult at design time to specify optimal threshold rules, it is even more challenging inferring precise performance models for the multitude of services. Recently, reinforcement learning have been applied to address this challenge. However, such approaches require many learning trials to stabilize at the beginning and when operational conditions vary thereby limiting their application under dynamic workloads. To this end, we extend the standard reinforcement learning approach in two ways: a) we formulate the system state as a fuzzy space and b) exploit a set of cooperative agents to explore multiple fuzzy states in parallel to speed up learning. Through multiple experiments on a real virtualized testbed, we demonstrate that our approach converges quickly, meets performance targets at high efficiency without explicit service models.

No abstract

No abstract