Section: New Results
Cloud applications and infrastructures
Participants : Frederico Alvares, Simon Dupont, Md Sabbir Hasan, Adrien Lebre, Thomas Ledoux, Jonathan Lejeune, Guillaume Le Louët, Jean-Marc Menaud, Jonathan Pastor, Mario Südholt.
In 2015, we have provided solutions for Cloud-based and distributed programming, virtual environments and data centers.
Cloud and distributed programming
Cloud Computing has provided important new means for the capacity management of resources. The elasticity and the economy of scale are the intrinsic elements that differentiate it from traditional computing paradigm.
A good capacity planning method is a necessary factor but not sufficient to fully exploit Cloud elasticity. In  , we propose innovative policies for resource management to achieve the optimal balance between capacity and quality of Cloud services The main idea is to finely control the scalability and the termination of virtual machines with respect to several criteria such as the lifecycle of the instances (e.g. initialization time) or their cost. The approach was evaluated on an Amazon EC2 cluster. Experimental results illustrate the soundness of the proposed approach and the impact of scalability/termination resource policies: a cost saving of as much as 30% can be achieved with a minimal number of violations, as small as 1%.
In order to improve Cloud elasticity, we advocate that the software layer can take part in the elasticity process as the overhead of software reconfiguration can be usually considered negligible compared to infrastructural costs. Thanks to this extra level of elasticity, we are able to define cloud reconfigurations that enact elasticity in both the software and infrastructure layers. In  , we present an autonomic approach to manage cloud elasticity in a cross-layered manner. First, we enhance cloud elasticity with the software elasticity model. Then, we describe how our autonomic cloud elasticity model relies on the dynamic selection of elasticity tactics. We present an experimental analysis of a subset of those elasticity tactics under different scenarios in order to provide insights on strategies that could drive the autonomic selection of the proper tactics to be applied.
Service-level agreement for the Cloud
Quality-of-service and SLA guarantees are among the major challenges of cloud-based services. In  , we first present a new cloud model called SLAaaS — SLA aware Service. SLAaaS considers QoS levels and SLA as first class citizens of cloud-based services. This model is orthogonal to other SaaS, PaaS, and IaaS cloud models, and may apply to any of them. More specifically, we make three contributions: (i) we provide a domain-specific language that allows to define SLA constraints in cloud services; (ii) we present a general control-theoretic approach for managing cloud service SLA; (iii) we apply our approach to MapReduce, locking, and e-commerce services.
Distributed multi-resource allocation
Generalized distributed mutual exclusion algorithms allow processes to concurrently access a set of shared resources. However, they must ensure an exclusive access to each resource. In order to avoid deadlocks, many of them are based on the strong assumption of a prior knowledge about conflicts between processes' requests. Some other approaches, which do not require such a knowledge, exploit broadcast mechanisms or a global lock, degrading message complexity and synchronization cost. We propose in   a new solution for shared resources allocation which reduces the communication between non-conflicting processes without a prior knowledge of processes conflicts. Performance evaluation results show that our solution improves resource use rate by a factor up to 20 compared to a global lock based algorithm.
Virtualization and data centers
In 2015, we have produced results and tools for the simulation of large-scale distributed algorithms, notably VM scheduling algorithms, have contributed new abstractions for storage systems and have devised new means for the introspection of Cloud infrastructures.
SimGrid / VMPlaceS
We have developed VMPlaceS  , a framework providing programming support for the definition of VM placement algorithms, execution support for their simulation at large scales, as well as new means for their trace-based analysis. VMPlaceS enables, in particular, the investigation of placement algorithms in the context of numerous and diverse real-world scenarios. To illustrate relevance of such a tool, we evaluated three different classes of virtualization environments: centralized, hierarchical and fully distributed placement algorithms. We showed that VMPlaceS facilitates the implementation and evaluation of variants of placement algorithms. The corresponding experiments have provided the first systematic results comparing these algorithms in environments including up to one thousand of nodes and ten thousands of VMs in most cases.
While such a number is already valuable and although we finalized the virtualization abstractions in SimGrid  , we are in touch with the core developers in order to improve the code of VMPlaceS with the ultimate objective of addressing infrastructures up to 100K physical machines and 1 Millions virtual machines over a period of one day.
The current version of VMPlaceS is available on a public git repository :http://beyondtheclouds.github.io/VMPlaceS/ .
Storage abstractions within the SimGrid framework
With the recent data deluge, storage is becoming the most important resource to master in modern computing infrastructures. Dimensioning and assessing the performance of storage systems are challenges for which simulation constitutes a sound approach. Unfortunately, only a few existing simulators of large scale distributed computing systems go beyond providing merely a notion of storage capacity. In 2015, we contributed to the SimGrid efforts toward the simulation of such systems  . Concretely, we characterized the performance behavior of several types of disks to derive a first model of storage resource. This model has been integrated within the SimGrid framework available under the LGPL license (http://simgrid.gforge.inria.fr ).
Cloud Computing has become a new technical and economic model for many IT companies. By virtualizing services, it allows for a more flexible management of datacenters capacities. However, its elasticity and its flexibility led to the explosion of virtual environments to manage. It’s common for a system administrator to manage several hundreds or thousands virtual machines. Without appropriate tools, this administration task may be impossible to achieve.
We purpose in  a decision support tool to detect virtual machines with atypical behavior. Virtual machines whose behavior is different from other VMs running in the data center are tagged as atypicals. Our analysis tool is based on a specific partitioning algorithm which identifies VM behaviors. This tool has been validated in production environments and is used by several companies.
To collect finer metrics (for security, energy management etc.), VM introspection an agent can be installed in a VM to intrusively supervise it or the hypervisor can be used to non-intrusively recover the introspection metrics. In the case of intrusive introspection, the agent installed on the VM operating system will retrieve a set of information related to the operating system operation. However, the installation of an agent in the virtual machine increases the cost of deploying the virtual machine and its resource consumption. The Virtual Machine Introspection (VMI) at the hypervisor level (non intrusively) offer a complete, consistent and untainted view of the VM state. This solution allows an isolation of the VMI mechanism from the guest OS, while allowing monitoring and modifying any state of the VM.
We have also provided a comprehensive summary on VM introspection techniques  . Existing VMI techniques are analyzed with respect to their approach to closing the "semantic gap" between the (low level) information provided by the hypervisor and the input to the security analysis.
Finally, we have introduced an extension to LibVmi to detect and monitor a process resource consumption inside a VM from the hypervisor  . This extension monitor process cpu and ram ressources without probe. This extension can detect abusive cpu resource usage and atypical ram utilization. This fine monitoring system can be used in many context (security, power consumption, fault tolerance).