Saving Energy in Grid Computing

Saving energy in Grid Computing. A. Fernández-Montes1, J. I. Sánchez-Venzalá1,J. A. Ortega1, L. González-Abril2 1Department of Computer Science, University of Sevilla, Sevilla, Spain {afdez,jisanchez,jortega}@us.es 2Department of Applied Economics, University of Sevilla, Sevilla, Spain [email protected] Abstract This article is focused on simulating the progress of Grid’5000 in order to choose best policies in terms of arranging jobs and managing resources. A Javabased simulator has been developed in order to replay the conditions of the Grid’5000 from recent years historical data . This way, an study of different policies has been carried out looking for energy efficiency. The policies studied are based in mathematical models which try to predict the most efficient behavior of the Grid’5000. 1 Introduction Saving energy is a key factor in computer science. Energy efficiency is looked for in all kind of systems from little devices to large scale computing. The huge amount of energy consumed by grid computing is a good reason to study saving energy methodologies either from an economical or ecological point of view. Grid operational policies must be mathematically analyzed in order to be optimized. The analysis that this paper presents has been accomplished over french Grid’5000, described next. 2 Grid’5000 Organization Grid’5000 is a scientific instrument designed to support experiment-driven research in all areas of computer science related to parallel, large-scale or distributed computing and networking. It aims to supply a highly reconfigurable, controlable and monitorable experimental platform to its users. The Grid’5000 provides a testbed which allows experiments in all the software layers between the network protocols up to the applications. Grid’5000 has been built upon a network of dedicated clusters. It is not an ad hoc grid. The infrastructure of Grid’5000 is geographically distributed on different sites, initially 9 in France: Bourdeaux, Grenoble, Lille, Lyon, Nancy, Orsay, Rennes, Sophia-Antipolis and Toulouse. Porto Alegre, in Brazil, and Luxemburg, are now officially becoming the 10th and 11th sites respectively. The project began in 2004 as an initiative of french ministry of Education and Research, INRIA, CNRS, the Universities of all sites and some regional councils. Figure 1: Grid’5000 french sites. The initial aim was to reach 5000 processors in the platform. It has been reframed at 5000 cores, and was reached during winter 2008-2009. On March 16th 2010, 1569 nodes (5808 cores) were in production in Grid’5000. Nowadays, sites see each others inside the same VLAN at 10Gbps thanks to the dark fiber infrastructure which connects them, in a not complete graph scheme. Grid’5000 allows experiments at grid or at cluster level, which guarantees a more homogeneous hardware and bandwidth, although grid level experiments are favored in planning. Each site of Grid’5000 hosts several clusters, because hardware has been acquired by incremental steps on each site, forming clusters at each purchase. Each cluster is formed by two kind of nodes: • Compute node, which conforms the base element of a cluster, on which computations are run. • Service node, which are dedicated to host the grid in- frastructure services, as control or deploy. Each node can supply several cores, which are the finest grain of resource in Grid’5000. 2.1 Tasks The platform can be used in two different modes: submissions and reservations. • Submission: an experiment is submitted and the scheduler decides when to run it. • Reservation: when a reservation of the platform for a certain time is made (although the experiment has to be launched interactively). The software used for task schedule is OAR. It is a resource manager (or batch scheduler) for large clusters which allows cluster users to submit or reserve nodes either in an interactive or in a batch mode. 3 Grid’5000 Simulator The Grid’5000 simulator tries to simulate the progress of the real Grid regarding jobs and resources operation. The objective is to be able to compute the energy consumed by Grid’5000 from historical data from past years, which is stored in a database, applying different policies of arranging jobs and managing resources. Arranging jobs policies are called Arranging Policies while managing resource policies are called Energy Policies. The simulator operation is based on an agenda where jobs are registered and a list of resources representing the real resources from the sites. The simulator starts to cover the agenda from the beginning to the end, modifying resources states as would be needed to execute them in the real world, taking in to account the policies established to manage resources and jobs. The consumed energy computation is made step by step by means of the information about energy consumption of each resource and resource states pointed out in the resource list. The result of simulation execution is a log where the behavior of grid, resources, and tasks according to the policies employed are shown together with the energy consumed computation. It has been implemented in Java, which makes possible an easy integration of new components, as the graphical interface, or the development of new extensions by others. 4 Energy Policies Energy policies establish the managing of the grid resources. They describe what to do with a resource when a task finishes its execution. There are several options: • Always On: it leaves resources always on, never switch them off. • Always Switch Off: it always switch resources off after a jobs execution. • Switch Off in Ts: after a jobs execution, it waits for a determined time (Ts) to switch-off the resource. Other energy policies are being studied and simulated in order to optimize energy saving in the grid. 5 Arranging Policies Arranging policies establish the arranging of the jobs for its execution. They can move a job from one resource to another, or even can move a planned job execution in time in order to taking advantages of resources that are already switched on. • Do Nothing: does not move jobs neither in time or from a resource to another, they are executed as they were defined in the agenda. • Simple Aggregation of Tasks: which tries to execute the jobs in the same resources, if possible, although it does not change planned jobs start time. Other arranging policies are being considered, in order to optimize the execution of tasks. 90