Tensor-Based Resource Utilization Characterization in a Large-Scale Cloud Infrastructure

Abstract

The introduction of virtualization and cloud computing has enabled a large number of containers/virtual machines to share computing resources. Nevertheless, the number and size of data centres are still on the rise, partly on account of an ever increasing amount of generated data and workloads worldwide. On the other hand, independent studies indicate that a large number of servers in contemporary data centres are underutilised. One of the strategies currently adopted by the research community in order to deal with resource inefficiency is dynamic workload consolidation. The idea behind is dynamically balancing the supply of computing, communication, and storage resources with the demand for resources. This entails populating physical servers with an optimal number of complementary workloads. Most existing or proposed approaches employ multi-variate optimisation to achieve this goal but do not easily lend themselves to fast and intuitive solutions. In this paper, we investigate the scope and usefulness of dimensionality reduction techniques (tensor decomposition) to identify execution and resource utilisation patterns in hosted containers/virtual machines. Our analysis is based on two large-scale data centres, one of them hosts 1190 commercial virtual machines on 59 physical computing servers and 29 physical storage servers organised in 9 clusters and the other 44373 containers on 3985 physical servers. Our analysis shows that spatial and temporal patters can be uncovered with tensor decomposition, based on which efficient clustering can be realised.