Evaluating performance degradation in NoSQL databases generated by virtualization

Abstract

There are many services in Cloud Systems currently on the market for hosting servers with different purposes. These environments are IT infrastructures frequently deployed in large data centers using Virtual Machines. Moreover, an increasing number of users are using massive online application resources with hosted databases on virtual machines. In light of this, there is a critical need to evaluate databases performance in virtual environments. This work aims to evaluate overhead generated by virtualization in two common NoSQL databases, Cassandra and MongoDB, by different virtualization techniques - full virtualization and paravirtualization. For this, we used the YCSB benchmark (Yahoo! Cloud Serving Benchmark) to drive performance tests, with the results being evaluated through statistical analysis. As findings, the performed experiments in both virtualization techniques demonstrated that the experimental design comprises most of the factors in MongoDB, different from Cassandra which comprises of just a few. In paravirtualization scenarios the environment factor was more sensitive to experimental variations overcoming factors such as threads and numbers of transaction. Both databases in full virtualization scenarios reached significant variation, however the threads and transactions factors were more significant.