Design and Evaluation of a Virtual Experimental Environment for Distributed Systems

Abstract

International audienceBetween simulation and experiments on real-scale testbeds, the combined use of emulation and virtualization provide a useful alternative for performing experiments on distributed systems such as clusters, grids, cloud computing or P2P systems. In this paper, we present Distem, a software tool to build distributed virtual experimental environments. Using an homogenenous set of nodes, Distem emulates a platform composed of heterogeneous nodes (in terms of number and performance of CPU cores), connected to a virtual network described using a realistic topology model. Distem relies on LXC, a low-overhead container-based virtualization solution, to achieve scalability and enable experiments with thousands of virtual nodes. Distem provides a set of user interfaces to accomodate different needs (command-line for interactive use, Ruby and REST APIs), is freely available and well documented. After a detailed description of Distem, we perform an experimental evaluation of several of its features.Entre la simulation et l'expérimentation sur des plates-formes réelles, l'usage combiné de l'émulation et de la virtualisation fournit une alternative utile pour réaliser des expériences sur des systèmes distribués tels que les clusters, grilles, le Cloud ou les systèmes P2P. Dans cet article, nous présentons Distem, un logiciel permettant de construire des environnements expérimentaux distribués virtuels. À partir d'un ensemble homogène de noeuds, Distem émule une plate-forme composée de noeuds hétérogènes (en termes de nombre et de performance de leurs coeurs CPU), connectés à un réseau virtuel décrit à partir d'un modèle de topologies réaliste. Distem se base sur LXC, une solution de virtualisation légère à base de conteneurs, pour obtenir des propriétés de passage à l'échelle satisfaisantes et permettre des expériences avec des milliers de noeuds virtuels. Distem fournit plusieurs interfaces utilisateurs permettant de s'adapter à différents besoins (ligne de commande pour l'usage interactif, Ruby, API REST), est librement disponible et bien documenté. Après une description détaillée de Distem, cet article présente une validation expérimentale de plusieurs de ses fonctionnalités

Distem: Evaluation of Fault Tolerance and Load Balancing Strategies in Real HPC Runtimes through Emulation

“…A framework for the evaluation of HPC applications 1) General overview of Distem: The features, general architecture and evaluation of Distem were presented in [4], but are shortly summarized here. Distem is able to emulate a heterogeneous virtual platform on top of a homogeneous cluster in order to meet the experimenter's requirements.…”

“…2) Emulate heterogeneous and evolving experimental conditions: We have previously shown [4] that Distem can be used to create virtual heterogeneous platforms from a homogeneous cluster, with vnodes with different frequencies and number of cores, different IO capabilities and network links with different bandwidth and latency. In the context of this work, we extended Distem so that those characteristics of the vnodes could be updated dynamically without redeploying the virtual platform.…”

Scalable and Reliable Data Broadcast with Kascade

“…To evaluate the fault-tolerance capabilities of Kascade described in Section III-D, we use the Distem [30] emulator. Distem is able to emulate a virtual platform over a regular cluster and is able to inject failures in the platform according to a specification provided by the user.…”

Emulation at Very Large Scale with Distem

“…Distem [1] is a distributed system emulator that leverages advanced Linux features like LXC, CPU frequency scaling and traffic control, to emulate a heterogeneous platform on top of a homogeneous cluster. Heterogeneity can be obtained by (1) specifying a virtual network topology where latency and bandwidth of the links can be defined, (2) emulating a degraded CPU capability, (3) executing several virtual nodes on a single physical node.…”

“…In previous work, we presented Distem [1], that provides an emulated platform on top of a regular cluster in order to facilitate the experimentation with distributed systems. In this paper, we focus on enhancing Distem to deal with large-scale experiments, typically those involving several dozens of thousands of nodes.…”