Section: New Results

Cloud programming and management

Cloud infrastructures

Our contributions regarding cloud infrastructures can be divided into three main topics described below: contributions related to (i) geo-distributed clouds (e.g., Fog and Edge computing), (ii) the convergence of Cloud and HPC infrastructures and (iii) the simulation of virtualized infrastructures.

Geo-distributed Clouds

Many academic and industry experts are now advocating a shift from large-centralized Cloud Computing infrastructures to massively small-geo-distributed data centers at the edge of the network. This new paradigm of utility computing is often called Fog and Edge Computing. Advantages of this paradigm are, among others, data-locality that enhances security aspects and response times for latency-critical applications, new energetic options because of reduced size of data centers (e.g., renewable energies), single point of failure avoidance etc. Among the obstacles to the adoption of this model though is the development of a convenient and powerful IaaS system capable of managing a significant number of remote data-centers in a unified way, including monitoring and data management issues in a decentralized environment.

In 2017, we achieved three main contributions toward this challenge.

In [12], we investigate how a holistic monitoring service for a Fog/Edge infrastructure, hosting next generation digital services, should be designed. Although several solutions have been proposed in the past for the monitoring of clusters, grids and cloud systems, none of those is well appropriate to the specific Fog and Edge Computing context. The contributions of this study are: (i) the problem statement, (ii) a classification and a qualitative analysis of major existing solutions, and (iii) a preliminary discussion of the impact of deployment strategies on the monitoring service.

In [6], [39], [17], we present successive studies related to the design and development of a first-class object store service for Fog/Edge facilities. After a deep analysis of major existing solutions (Ceph, Cassandra ...), we designed a proposal that combines Scale-out Network Attached Storage systems (NAS) and IPFS, a BitTorrent-based object store spread throughout the Fog/Edge infrastructure. Without impacting the IPFS advantages particularly in terms of data mobility, the use of a Scale-out NAS on each site reduces the inter-site exchanges that are costly but mandatory for the metadata management in the original IPFS implementation. Several experiments conducted on Grid'5000 testbed are analysed and corroborate, first, the benefit of using an object store service spread at the Edge, and second, the importance of mitigating inter-site accesses. Ongoing activities are related to the management of meta data information in order to benefit from data movements.

Finally, in [26], we introduce the premises of a fog/edge resource management system by leveraging the OpenStack software, a leading IaaS manager in the industry. The novelty of the presented prototype is to operate such an Internet-scale IaaS platform in a fully decentralized manner, using P2P mechanisms to achieve high flexibility and avoid single points of failure. More precisely, we revised the OpenStack Nova service (i.e., virtual machine management and allocation) by leveraging a distributed key/value store instead of the centralized SQL backend. We present experiments that validate the correct behavior and gives performance trends of our prototype through an emulation of several data-centers using Grid'5000 testbed.

Cloud and HPC convergence

Geo-distribution of Cloud Infrastructures is not the only current trend of utility computing. Another important challenge is to reach the convergence of Cloud and HPC infrastructures, in other words on-demand HPC. Among challenges of this convergence is, for example, the enhancement of the use of light virtualization techniques on HPC systems, as well as the enhancement of mechanisms to be able to consolidate those VMs without deteriorating the performance of HPC applications, thus minimizing interferences between applications.

In [36], we present Eley, a burst buffer solution that helps to accelerate the performance of Big Data applications while guaranteeing the QoS of HPC applications. To achieve this goal, Eley embraces interference-aware prefetching technique that makes reading data input faster while introducing low interference for HPC applications. Specifically, we equip the prefetcher with five optimization actions including No Action, Full Delay, Partial Delay, Scale Up and Scale Down. It iteratively chooses the best action to optimize the prefetching while guaranteeing the pre-defined QoS requirement of HPC applications (i.e., the deadline constraint for the completion of each I/O phase). Evaluations using a wide range of Big Data and HPC applications show the effectiveness of Eley in reducing the execution time of Big Data applications (shorter map phase) while maintaining the QoS of HPC applications.

Virtualization simulation

Finally, it is important to be able to simulate the behavior of proposals for the future architectures. However, current models for virtualized resources are not accurate.

In [32], we present our latest results regarding virtualization abstractions and models for cloud simulation toolkits. Cloud simulators still do not provide accurate models for most Virtual Machine (VM) operations. This leads to incorrect results in evaluating real cloud systems. Following previous works on live-migration, we discuss an experimental study we conducted in order to propose a first-class VM boot time model. Most cloud simulators often ignore the VM boot time or give a naive model to represent it. After studying the relationship between the VM boot time and different system parameters such as CPU utilization, memory usage, I/O and network bandwidth, we introduce a first boot time model that could be integrated into current cloud simulators. Through experiments, we also show that our model correctly reproduced the boot time of a VM under different resources contention.

Deployment and reconfiguration in the Cloud

Being able to manage the new generation of utility computing infrastructures is an important step to build useful system building blocks. The next step is to be able to perform initial deployment of any kind of distributed software (i.e., systems, frameworks or applications) on those infrastructures, thus dealing with a complex process that includes interactions between building blocks such as virtual machine management, optimized deployment plans, monitoring of deployment etc. Such deployment processes cannot be handled manually anymore, for this reason automatic deployments tools have to be designed according to the challenges of new infrastructures (e.g., geo-distribution, hybrid infrastructures etc.). Moreover, as distributed software are more and more dynamic (i.e., reconfiguring themselves at runtime), reconfiguration and self-management capabilities should be handled in an efficient and scalable manner.

Initial deployment and placement strategies

When focusing on the initial deployment, many challenges should already need to be addressed such as placement of distributed software onto virtual machines, themselves being placed onto physical resources. This kind of placement problem can be modeled in many different ways, such as linear or constraint programming or graph partitioning. Most of the time a multi-objective NP-hard problem is formulated, and specific heuristics have to be built to reach scalable solutions.

In [18], we present new specific placement constraints and objectives adapted to hybrid clouds infrastructures, and we address this problem through constraint programming. Furthermore we evaluate the expressivity and performance of the solution on a real case study. In the Cloud, if public providers enable simple access to resources for companies and users who have sporadic computation or storage needs, private clouds could sometimes be preferred for security or privacy reasons, or for cost reasons due to a high frequency usage of services. However, in many cases a choice between public or private clouds does not fulfill all requirements of companies and hybrid cloud infrastructures should be preferred. Solutions have already been proposed to address hybrid cloud infrastructures, however most of the time the placement of a distributed software on such infrastructure has to be indicated manually.

In [37], we present a geo-aware graph partitioning method named G-Cut, which aims at minimizing the inter-DC data transfer time of graph processing jobs in geo-distributed DCs while satisfying the WAN usage budget. G-Cut adopts two novel optimization phases which address the two challenges in WAN usage and network heterogeneities separately. G-Cut can be also applied to partition dynamic graphs thanks to its light-weight runtime overhead. We evaluate the effectiveness and efficiency of G-Cut using real-world graphs with both real geo-distributed DCs and simulations. Evaluation results demonstrate that effectiveness of G-Cut in reducing the inter-DC data transfer time and the WAN usage with a low runtime overhead.

Many other challenges than placement rise from the initial deployment. In [20], we present a survey of existing deployment tools that have been used in production to deploy OpenStack, which is a complex distributed system composed of more than a hundred different services. To fully understand how IaaSes are deployed today, we propose in this paper an overall model of the application deployment process that describes each step with their interactions. This model then serves as the basis to analyse five different deployment tools used to deploy OpenStack in production: Kolla, Enos, Juju, Kubernetes, and TripleO. Finally, a comparison is provided and the results are discussed to extend this analysis.

Capacity planning and scheduling

While a placement problem is a discrete problem at a given instant, some other challenges of deployment and reconfiguration may include the time dimension leading to scheduling optimization.

in [30] we have proposed two original workload prediction models for Cloud infrastructures. These two models, respectively based on constraint programming and neural networks, focus on predicting the CPU usage of physical servers in a Cloud data center. The predictions could then be exploited for designing energy-efficient resource allocation mechanisms like scheduling heuristics or over-commitment policies. We also provide an efficient trace generator based on constraint satisfaction problem and using a small amount of real traces. Such a generator can overcome availability issues of extensive real workload traces employed for optimization heuristics validation. While neural networks exhibit higher prediction capabilities, constraint programming techniques are more suitable for trace generation, thus making both techniques complementary.

Reconfiguration and self-management

Being able to handle the dynamicity of hardware, system building blocks, middleware and applications is a great challenge of today's and future utility computing systems. On large infrastructures such as Cloud, Fog or Edge Computing, manual administration of such dynamicity is not feasible. The automatic management of reconfiguration, or self-management of software is of great importance to guarantee reliability, fault tolerance, security, and cost and energy optimization.

In [4], in order to improve the self-adaptive behaviors in the context of Component-based Architecture, we design self-adaptive software components based on logical discrete control approaches, in which the self-adaptive behavioural models enrich component controllers with a knowledge not only on events, configurations and past history, but also with possible future configurations. This article provides the description, implementation and discussion of Ctrl-F, a Domain-specific Language whose objective is to provide high-level support for describing these control policies. In [13], we extended Ctrl-F with modularity capabilities. Apart from the benefits of reuse and substitutability of Ctrl-F programs, modularity allows to break down the combinatorial explosion intrinsic to the generation of correct-by-construction controllers in the compilation process of Ctrl-F. A further advantage of modularity is that the executable code, that is, the controllers resulting from that compilation, are loss-coupled and can therefore be deployed and executed in a distributed fashion.

However, higher abstraction-level tools also have to be proposed for reconfiguration. In [21], we introduce ElaScript, a Domain Specific Language (DSL) which offers Cloud administrators a simple and concise way to define complex elasticity-based reconfiguration plans. ElaScript is capable of dealing with both infrastructure and software elasticities, independently or together, in a coordinated way. We validate our approach by first showing the interest to have a DSL offering multiple levels of control for Cloud elasticity, and then by showing its integration with a realistic well-known application benchmark deployed in OpenStack and the Grid’5000 infrastructure testbed.

Finally, self-management can be applied at many different levels of the Cloud paradigm, from infrastructure reconfigurations to application topology reconfigurations. In practice these reconfiguration mechanisms are tightly coupled. For example, a change in the infrastructure could lead to the re-deployment of virtual machines upon it that could lead itself to application reconfigurations. In [27], we advocate that Cloud services, regardless of the layer, may share the same consumer/provider-based abstract model. From that model, we can derive a unique and generic Autonomic Manager (AM) that can be used to manage any XaaS (Everything-as-a-Service) layer defined with that model. The paper proposes such an abstract (although extensible) model along with a generic constraint-based AM that reasons on abstract concepts, service dependencies as well as SLA (Service Level Agreements) constraints in order to find the optimal configuration for the modeled XaaS. The genericity of our approach are shown and discussed through two motivating examples and a qualitative experiment has been carried out in order to show the applicability of our approach as well as to discuss its limitations.