| dc.description.abstract | With the rise of cloud computing, data centers have been called to play a main role in the Internet scenario nowadays. Despite this relevance, they are probably far from their zenith yet due to the ever increasing demand of contents to be stored in and distributed by the cloud, the need of computing power or the larger and larger amounts of data being analyzed by top companies such as Google, Microsoft or Amazon.
However, everything is not always a bed of roses. Having a data center entails two major issues: they are terribly expensive to build, and they consume huge amounts of power being, therefore, terribly expensive to maintain. For this reason, cutting down the cost of building and increasing the energy efficiency (and hence reducing the carbon footprint) of data centers has been one of the hottest research topics during the last years. In this thesis we propose different techniques that can have an impact in both the building and the maintenance costs of data centers
of any size, from small scale to large flagship data centers.
The first part of the thesis is devoted to structural issues. We start by analyzing the bisection (band)width of a topology, of product graphs in particular, a useful parameter to compare and choose among different data center topologies. In that same part we describe the problem of
deploying the servers in a data center as a Multidimensional Arrangement Problem (MAP) and propose a heuristic to reduce the deployment and wiring costs.
We target energy efficiency in data centers in the second part of the thesis. We first propose a method to reduce the energy consumption in the data center network: rate adaptation. Rate adaptation is based on the idea of energy proportionality and aims to consume power on network
devices proportionally to the load on their links. Our analysis proves that just using rate adaptation we may achieve average energy savings in the order of a 30-40% and up to a 60% depending on the network topology.
We continue by characterizing the power requirements of a data center server given that, in order to properly increase the energy efficiency of a data center, we first need to understand how energy is being consumed. We present an exhaustive empirical characterization of the power
requirements of multiple components of data center servers, namely, the CPU, the disks, and the network card. To do so, we devise different experiments to stress these components, taking into account the multiple available frequencies as well as the fact that we are working with multicore servers. In these experiments, we measure their energy consumption and identify their optimal operational points. Our study proves that the curve that defines the minimal power consumption of the CPU, as a function of the load in Active Cycles Per Second (ACPS), is neither concave
nor purely convex. Moreover, it definitively has a superlinear dependence on the load. We also validate the accuracy of the model derived from our characterization by running different Hadoop applications in diverse scenarios obtaining an error below 4:1% on average.
The last topic we study is the Virtual Machine Assignment problem (VMA), i.e., optimizing how virtual machines (VMs) are assigned to physical machines (PMs) in data centers. Our
optimization target is to minimize the power consumed by all the PMs when considering that power consumption depends superlinearly on the load. We study four different VMA problems, depending on whether the number of PMs and their capacity are bounded or not. We study their complexity and perform an offline and online analysis of these problems. The online analysis is complemented with simulations that show that the online algorithms we propose consume substantially less power than other state of the art assignment algorithms. | |
