Implantação de bancos de dados distribuídos em um cluster de baixo consumo

Abstract

The constant technological advances, Web technologies, mobile devices and the popularization of the IoT (Internet of Things), have caused an exponential growth in the volume of data never seen before. This scenario, although positive from a technological point of view, brings many challenges to data processing and storage centers, making necessary the adoption of tools and technologies most adapted to deal with this paradigm. Thus, distributed databases are the most suitable solution for this scenario, since characteristics such as horizontal scalability, elasticity and high availability allow such technologies to keep up with the growing data volume and sources. However, in the same way that the volume of data increases, also increases the demands of computational power, investment, space and energy consumption of the infrastructures as a whole to provide the environment of the distributed databases. In that way, this work explores the use of a low-power cluster composed of SBCs (Single Board Computers), for the implementation of distributed databases, in order to validate the feasibility of using this type of cluster as a compact, cheap and with lower energy consumption alternative for the common data center infrastructures. In that way, this work explores the use of a low-power cluster composed of SBCs (Single Board Computers), for the implementation of distributed databases, in order to validate the feasibility of using this type of cluster as a compact, cheap and with lower energy consumption alternative for the common data center infrastructures. Fifteen Raspberry Pi 3 B devices were used to compose the cluster, which supports a virtualization layer of Docker containers orchestrated by the Docker Swarm tool. The performance of Cassandra, Hbase and PostgreSQL/Citus databases on the SBC cluster was evaluated, using YCSB benchmark workloads to analyze execution time, latency and throughput in scenarios with different replication factors. The results show that, in general, Cassandra outformed the other databases and obtained the best results, showing no influence by the replication factor. The results for Hbase and Citus were heavily penalized by the increase of the replication factor. Furthermore, the results also prove the ability of the low-power cluster to meet the requirements of distributed systems used in real scenarios, allowing to the distributed environment elasticity and high availability.