Section: New Results

Cloud Computing, Virtualization and Data Centers

Participants : Frederico Alvares, Gustavo Bervian Brand, Thomas Chavrier, Fabien Hermenier, Adrien Lèbre, Thomas Ledoux, Guillaume Le Louët, Jean-Marc Menaud, Hien Nguyen Van, Rémy Pottier, Flavien Quesnel.

In the context of Cloud computing ASCOLA members have principally worked this year on capacity planning solutions for large scale distributed system. Capacity planning is the process of planning, analyzing, sizing, managing and optimizing capacity to satisfy demand in a timely manner and at a reasonable cost.

Applied to distributed systems like the Cloud, a capacity planning solution must mainly provide necessary resources for the proper execution of applications and respect service-level agreements in a large distributed environment.

The main challenges in this context are: scalability, fault tolerance and reactivity of the solution in a large-scale distributed system; analyzing, sizing, and optimizing resources to minimize the cost (energy, human, hardware etc.); and profiling, adapting application to ensure the quality of service (throughput, response time, availability etc.).

Our solutions are mainly based on virtualized infrastructures. Our main results concern the management and the execution of applications by leveraging virtualization capabilities on cloud infrastructures and the investigation of solutions that aim at optimizing the trade-off between performance and energy costs of both applications and Cloud resources.

Virtualization and Job Management

This year, in cooperation with the Myriads project-team from INRIA Rennes-Bretagne Atlantique, we have continued to address resource-management issues concerning the federation of very large scale platforms. We have completed our approach aiming at the automatic adaptation of both hardware and software resources to the needs of the applications through a unique method. For each application, scientists describe the requirements in terms of both hardware and software expectations through the definition of a Virtual Platform (VP) and a Virtual System Environment (VSE) [18]

In addition, we started to address the design and the implementation of a fully distributed cloud OS. Designing and implementing such models is a tedious and complex task. However, as well as research studies on OSes and hypervisors are complementary at the node level, we showed that virtualization frameworks can benefit from lessons learned from distributed operating system proposals [34] . Leveraging this preliminary result, we designed and developed a first proof-of-concept of a fully distributed scheduler [33] . This system makes it possible to schedule VMs cooperatively and dynamically in large scale distributed systems. Preliminary results showed that our scheduler was more reactive. This building block is a first element of a more complete cloud OS, entitled DISCOVERY (DIStributed and COoperative mechanisms to manage Virtual EnviRonments autonomicallY) [30] . The ultimate goal of this system is to overcome the main limitations of the traditional server-centric solutions. The system, currently under development, relies on a peer-to-peer model where each agent can efficiently deploy, dynamically schedule and periodically checkpoint the virtual environments s/he manage.

We have contributed new results on the Entropy software and extended our solution VM that features dynamic consolidation. In [44] and [38] we extended our dynamic consolidation manager to take into account not only resource constraints but also the placement constraints of highly available (HA) applications. In fact, most previous dynamic consolidation systems optimize the placement of the VMs according to their resource usage but do not consider the application placement constraints that are required to achieve both high availability and scalability. Our approach provides efficient dynamic consolidation while guaranteeing to the application administrator that placement requirements will be satisfied and relieving the data center administrator of the burden of considering the constraints of the applications when performing maintenance.

In the same domain, Jean-Marc Menaud has defended his habilitation (HdR - Habilitation à diriger des recherches) [13] on Jun. 22, 2011. One part of this HDR focuses on dynamic adaptation strategies for cluster administration. We have proposed a dedicated language for virtual machines management and one particular feature of our solution is to use a constraint solver to provide an appropriate placement. Based on these results, our recent contributions address the problem of data center energy consumption.

Moreover, we have continued to analyze how energy concerns can be addressed in large scale distributed infrastructures.

Optimization of Energy Consumption in Data Centers

As a direct consequence of the increasing popularity of Cloud Computing solutions, data centers are growing at a fast rate and hence have to face difficult energy consumption issue. Available solutions rely on Cloud Computing models and virtualization techniques to scale up/down applications based on their performance metrics. Although those proposals can reduce the energy footprint of applications and, by transitivity, of cloud infrastructures, they do not consider the internal characteristics of applications to finely define a trade-off between QoS properties of applications and their energy footprint. In [22] , we propose a self-adaptation approach that considers both application internals and system properties to reduce the energy footprint in cloud infrastructures. Each application and the infrastructure are equipped with their own control loop, which allows them to autonomously optimize their executions. Simulations show that the approach may lead to appreciable energy savings without interfering on application provider revenues.

Cloud Computing and SLA Management

Cloud computing is a paradigm for enabling remote, on-demand access to a set of configurable computing resources as a service. The pay-per-use model enables service providers to offer their services to customers at different Quality-of-Service (QoS) levels. These QoS parameters are used to compose service-level agreements (SLAs) between a service provider and a service consumer. A main challenge for a service provider is to manage SLAs for its service consumers, i.e., automatically determine the appropriate resources required from the lower layer in order to respect the QoS requirements of the consumers. In [27] , we have proposed an optimization framework driven by consumer preferences to address the SLA dependencies problem across the different cloud layers as well as the need of flexibility and dynamicity required by the domain of Cloud computing. Our approach aims at selecting the optimal vertical business process designed using cross-layer cloud services and enforcing SLA dependencies between layers. Experimental results demonstrate the flexibility and effectiveness of our approach.