Arquitetura de um cluster computacional de baixo consumo e com proporcionalidade energética

Abstract

One of the main challenges for the so-called Green Computing is to get a better relation between the amount of work performed by the computational infrastructure and the energy consumption to maintain it, providing better energy efficiency. This work presents the architecture of a computing cluster with low energy consumption that powers on or off a number of running machines automatically and dynamically. The quantity of enabled devices adjusts according to the actual processing demand, which avoids unnecessarily powered equipment and increases the overall system power efficiency. In order to carry out its development, a theoretical model is proposed, discussed, and implemented through the NPi-Cluster, a cluster composed of Raspberry Pi devices. To prove the proposed model feasibility, NPi-Cluster is used as a web server with load balancing. Data gathered shows that NPi-Cluster has adequate performance when compared to other web servers running on traditional server architectures, however with less power consumption. A 7-machine cluster running at maximum performance is able to handle more than 450 simultaneous requests, with about 1000 transactions per second. The power consumption required to do it is about 15 Watts, which is equivalent to a energy-saving light bulb or a computer in suspended mode that does not perform any task. When the requests demand is low, the power consumption is dynamically reduced until less than 2 Watts. Besides to being able to handle workloads with acceptable quality of service, the proposed cluster also provides high availability by avoiding single points of failure.