Section: New Results

Cloud Resource Management

Participants : Eddy Caron, Hadrien Croubois, Jad Darrous, Christian Perez.

Nitro: Network-Aware Virtual Machine Image Management in Geo-Distributed Clouds

Recently, most large cloud providers, like Amazon and Microsoft, replicate their Virtual Machine Images (VMIs) on multiple geographically distributed data centers to offer fast service provisioning. Provisioning a service may require to transfer a VMI over the wide-area network (WAN) and therefore is dictated by the distribution of VMIs and the network bandwidth in-between sites. Nevertheless, existing methods to facilitate VMI management (i.e., retrieving VMIs) overlook network heterogeneity in geo-distributed clouds. In [19], we design, implement and evaluate Nitro, a novel VMI management system that helps to minimize the transfer time of VMIs over a heterogeneous WAN. To achieve this goal, Nitro incorporates two complementary features. First, it makes use of deduplication to reduce the amount of data which will be transferred due to the high similarities within an image and in-between images. Second, Nitro is equipped with a network-aware data transfer strategy to effectively exploit links with high bandwidth when acquiring data and thus expedites the provisioning time. Experimental results show that our network-aware data transfer strategy offers the optimal solution when acquiring VMIs while introducing minimal overhead. Moreover, Nitro outperforms state-of-the-art VMI storage systems (e.g., OpenStack Swift) by up to 77%.

Toward an Autonomic Engine for Scientific Workflows and Elastic Cloud Infrastructure

The constant development of scientific and industrial computation infrastructures requires the concurrent development of scheduling and deployment mechanisms to manage such infrastructures. Throughout the last decade, the emergence of the Cloud paradigm raised many hopes, but achieving full platform autonomicity is still an ongoing challenge. We built a workflow engine that integrated the logic needed to manage workflow execution and Cloud deployment on its own. More precisely, we focus on Cloud solutions with a dedicated Data as a Service (DaaS) data management component. Our objective was to automate the execution of workflows submitted by many users on elastic Cloud resources. This contribution proposes a modular middleware infrastructure and details the implementation of the underlying modules:

• A workflow clustering algorithm that optimises data locality in the context of DaaS-centered communications;

• A dynamic scheduler that executes clustered workflows on Cloud resources;

• A deployment manager that handles the allocation and deallocation of Cloud resources according to the workload characteristics and users’ requirements.

All these modules have been implemented in a simulator to analyse their behaviour and measure their effectiveness when running both synthetic and real scientific workflows. We also implemented these modules in the Diet middleware to give it new features and prove the versatility of this approach. Simulation running the WASABI workflow (waves analysis based inference, a framework for the reconstruction of gene regulatory networks) showed that our approach can decrease the deployment cost by up to 44% while meeting the required deadlines [13].

Madeus: A Formal Deployment Model

Distributed software architecture is composed of multiple interacting modules, or components. Deploying such software consists in installing them on a given infrastructure and leading them to a functional state. However, since each module has its own life cycle and might have various dependencies with other modules, deploying such software is a very tedious task, particularly on massively distributed and heterogeneous infrastructures. To address this problem, many solutions have been designed to automate the deployment process. In [18], we introduce Madeus, a component-based deployment model for complex distributed software. Madeus accurately describes the life cycle of each component by a Petri net structure, and is able to finely express the dependencies between components. The overall dependency graph it produces is then used to reduce deployment time by parallelizing deployment actions. While this increases the precision and performance of the model, it also increases its complexity. For this reason, the operational semantics need to be clearly defined to prove results such as the termination of a deployment. In this paper, we formally describe the operational semantics of Madeus, and show how it can be used in a use-case: the deployment of a real and large distributed software (i.e., , OpenStack).

In [18], we have proposed an extension based on component behavioral interfaces to the Aeolus component model to better separate the concerns of component users (e.g., application architect) from component developers.