Keywords: energy, heuristic, virtual machines, grid.

power consumption as a result of computational servers and data centers. We will consider the opposite problem of load balancing: to concentrate the load on a minimum number of machines. Some solutions, algorithms, include numerous approaches, through the network protocols [1], modeling of consumption [2], scheduling of tasks by predicting idle times [3]. To evaluate improvements we choose to use a grid simulator. We present here a series of experiments on the anti-load balancing for significantly reducing energy consumption.

To analyse the results of our simulations, we focus on the energy of the system. Indeed, the energy of the system will direct a guide to assess the impact of our method on the reduction of energy. We first present the evolution of the energy of a system to observe how it behaves when we vary the number of guests. In the first series of experiments we first run two jobs on two machines, then on two different machines. In the second experiment we vary the load of resources by performing a task migration. VMs on hosts that are at minimum load (in our case a load less than 50%) are selected as potential node to migrate their load.

In the second experiment we check the evolution of energy in the execution of jobs with no migration and with migration. We note that the evolution of the gap as the number of jobs increases. As might be expected, where the number of jobs is high, around 103, the energy is the same for the two versions. This is due to saturation of the data center, when migration can no longer be done.

For the third experiment, we try to avoid having a heating point. With a low load, energy consumption rises as the number of jobs increases. However, energy consuption decreases if nodes are not overloaded, whatever the load is. With the last experiment we study three different aggressive modes:

• Succ curve: the least loaded host transfers its load to the least loaded node.
• Dec curve: we compare the least load node to other nodes starting with the most loaded node. Jobs are migrated to the first node that can receive them.
• Rand curve: the load of the node is migrated to a randomly chosen node

It is clear that the random choice of the node that receives the load increases the electric consumption.

The use of anti-load balancing provides a power gain in managing data centers. Our solution is based on a centralized architecture. The results of our simulations show that our solution fits better when the grid size increases. A logical extension of this work, to better optimize our approach is to integrate quality of service (QoS) but also to address the distributed solution.

1

B. Rountree, D.K. Lowenthal, S. Funk, V.W. Freeh, B.R. de Supinski, M. Schulz, "Bounding energy consumption in large-scale MPI programs", Proceedings of the ACM/IEEE Conference on High Performance Networking and Computing, SC 2007, Reno, Nevada, USA, November 10-16, 2007. doi:10.1145/1362622.1362688

2

H. Hlavacs, G. Da Costa, J.M. Pierson, "Energy Consumption of Residential and Professional Switches", Rapport de recherche, IRIT, Université Paul Sabatier, Toulouse, 2009.

3

G. Da-Costa, J.P. Gelas, Y. Georgiou, L. Lefévre, A.C. Orgerie, J.M. Pierson, O. Richard, K. Sharma, "The GREEN-NET Framework : Energy Efficiency in Large Scale Distributed Systems", HPPAC 2009: High Performance Power Aware Computing Workshop in conjunction with IPDPS 2009. doi:10.1109/IPDPS.2009.5160975

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £85 +P&P)