Section: New Results

Resource management in Cloud computing

Participants : Frederico Alvares, Gustavo Bervian Brand, Yousri Kouki, Adrien Lèbre, Thomas Ledoux, Guillaume Le Louët, Jean-Marc Menaud, Jonathan Pastor, Flavien Quesnel, Mario Südholt.

We have contributed on several topics: multiple autonomic managers for Cloud infrastructure, SLA management for Cloud elasticity, fully distributed and autonomous virtual machine scheduling, and simulator toolkits for IaaS platforms.

Cloud infrastructure based on multiple autonomic managers

One of the main reasons for the wide adoption of Cloud Computing is the concept of elasticity. Implementing elasticity to tackle varying workloads while optimizing infrastructures (e.g. utilization rate) and fulfilling the application requirements on Quality of Service should be addressed by self-adaptation techniques able to manage complexity and dynamism. However, since Cloud systems are organized in different but dependent Cloud layers, self-management decisions taken in isolation in a certain layer may indirectly interfere with the decision taken by an other layer. Indeed, non-coordinated managers may lead to conflicting decisions and consequently to non-desired states.

We have proposed a framework for the coordination of multiple autonomic managers in cloud environments [25] . The PhD thesis of Frederico Alvares [12] , defended in April 2013, is based on this framework. This thesis proposes a self-adaptation approach that considers both application internals (architectural elasticity) and infrastructure (resource elasticity), managed by multiple autonomic managers, to reduce the energy footprint in Cloud infrastructures.

SLA Management for Cloud elasticity

Elasticity is the intrinsic element that differentiates Cloud Computing from traditional computing paradigms, since it allows service providers to rapidly adjust their needs for resources to absorb the demand and hence guarantee a minimum level of Quality of Service (QoS) that respects the Service Level Agreements (SLAs) previously defined with their clients. However, due to non-negligible resource initiation time, network fluctuations or unpredictable workload, it becomes hard to guarantee QoS levels and SLA violations may occur. The main challenge of service providers is to maintain its consumer's satisfaction while minimizing the service costs due to resources fees. The PhD thesis of Yousri Kouki [13] , defended in December, proposes different contributions to address this issue: CSLA, a specific language to describe SLA for Cloud services ; HybridScale, an auto-scaling framework driven by SLA [39] , [17] .

Fully Distributed and Autonomous Virtualized Environments

We have consolidated the DVMS system to obtain a fully distributed virtual machine scheduler [44] . This system makes it possible to schedule VMs cooperatively and dynamically in large scale distributed systems. Simulations (up to 64K VMs) and real experiments both conducted on the Grid'5000 large-scale distributed system [44] showed that DVMS is scalable. This building block is a first element of a more complete cloud OS, entitled DISCOVERY (DIStributed and COoperative mechanisms to manage Virtual EnviRonments autonomicallY) [56] . The ultimate goal of this system is to overcome the main limitations of the traditional server-centric solutions. The system, currently under investigation in the context of the Jonathan Pastor's PhD, relies on a peer-to-peer model where each agent can efficiently deploy, dynamically schedule and periodically checkpoint the virtual environments it manages.

Testing the cloud

Computer science, as other sciences, needs instruments to validate theoretical research results, as well as software developments. Although simulation and emulation are generally used to get a glance of the behavior of new algorithms, they use over-simplified models in order to reduce their execution time and thus cannot be accurate enough. Leveraging a scientific instrument to perform actual experiments is an undeniable advantage. However, conducting experiments on real environments is still too often a challenge for researchers, students, and practitioners: first, because of the unavailability of dedicated resources, and second, because of the inability to create controlled experimental conditions, and to deal with the wide variability of software requirements. During 2013, we have contributed to a new topic addressing the “testing the cloud” challenge. First, we have presented the latest mechanisms we have designed to enable the automated deployment of the major open-source IaaS cloudkits (i.e., Nimbus, OpenNebula, CloudStack, and OpenStack) on Grid’5000 [26] . Providing automatic, isolated and reproducible deployments of cloud environments lets end-users study and compare each solution or simply leverage one of them to perform higher-level cloud experiments (such as investigating Map/Reduce frameworks or applications). Moreover, we have presented EXECO, a library that provides easy and efficient control of large scale experiments through a set of tools well as tools designed for scripting distributed computing experiments on any computing platform. We have illustrated its interest by presenting two experiments dealing with virtualization technologies on the Grid’5000 testbed [37] .

Adding virtualization abstractions into the Simgrid toolkit

In the context of the ANR SONGS project and in collaboration with Takahiro Hirofuchi, researcher at AIST (Japan), we have extended the Simgrid framework to be able to simulate virtualized distributed infrastructures [35] . In addition, we have proposed the first class support of live migration operations within such a simulator toolkit for large scale distributed infrastructures. We have developed a resource share calculation mechanism for VMs and a live migration model implementing the precopy migration algorithm of Qemu/KVM. We have confirmed that our simulation framework correctly reproduced live migration behaviors of the real world under various conditions [36] .

Power and energy management in the cloud

Power management has become one of the main challenges for data center infrastructures. Currently, the cost of powering a server is approaching the cost of the server hardware itself, and, in a near future, the former will continue to increase, while the latter will go down. In this context, virtualization is used to decrease the number of servers, and increase the efficiency of the remaining ones.

First, in [43] we have proposed an approach and a model to estimate the total power consumption of a virtual machine, by taking into account its static (e.g. memory) and dynamic (e.g. CPU) consumption of resources. Second, we have rewritten the Entropy framework (in OptiPlace) to give it the support of external models, named views. Entropy, based on the Constraint Programming solver Choco written in Java, does not really scale well. We have studied Entropy's scalability properties [32] and have then integrated heuristics and constraints in OptiPlace [40] .

The evaluation of these policies on real infrastructures has become an important and difficult issue. The corresponding techniques have become so complex that there is a need for load injection frameworks able to inject resource load in a tested datacenter instead of model-driven simulation. For this reason we have developed StressCloud [41] , [51] , a framework to manipulate the activities of a group of Virtual Machines and observe the resulting performance.