Energy Efficiency of Data Storage Systems in Cluster Computing
Tuominen, Mikko Rikhartti · Tampere University Institutional Repository (Tampere University) · 2011
Energy efficiency is an important part of the development of any technology. Cluster computing is no exception. As the energy prices rise, the costs of running a cluster can easily overcome the costs of buying one. A euro saved is a euro earned. This thesis examines and compares different hardware level approaches and software level configurations used in clusters to storage data. Solid state drives are not commonly used in clusters and one of the goals of this thesis is to study whether or not this relatively new technology is suitable to be used in clusters. The main goal is to understand what affects to the energy efficiency of a cluster from a data storage point of view. To reach these goals, the performance and energy consumption of a cluster, with different system configurations, is measured and analysed. These results can further be used to optimise existing clusters. The thesis is divided into two parts. In the literature study part, issues related to energy efficiency, data storage models, block devices, file systems and I/O schedulers are studied. In the experimental part, the test environment is introduced in detail and the results are reported and analysed. The tests are conducted using the CMS software with real LHC data to simulate heavy physics computing. During these tests, both hard disk and solid state drives are used with three different data storage schemes; a distributed approach with GlusterFS (a distributed file system) on compute nodes, a centralised approach with dedicated file server and a local approach with drives in the compute nodes of the cluster. The test results reveal that no significant gain is achieved by using solid state drives. Another key result is that a cluster can suffer from a major performance loss if the file system and I/O scheduler is not properly selected. The conclusion of this thesis is, that although there is no fundamental reason why solid state drives should not be used in clusters, considering the multifold price and low capacity compared to hard disk drives, it is not justifiable. As the development of solid state drives progress, a new study is in order. If the prices decline and storage capacity increases, solid state drives could abolish mechanical drives. /Kir11