Diagnosing, predicting and managing application performance in virtualised multi-tenant clouds

As the computing industry enters the cloud era, multicore architectures and virtualisation technologies are replacing traditional IT infrastructures for several reasons including reduced infrastructure costs, lower energy consumption and ease of management. Cloud-based software systems are expected to deliver reliable performance under dynamic workloads while efficiently allocating resources. However, with the increasing diversity and sophistication of the environment, managing performance of applications in such environments becomes difficult. The primary goal of this thesis is to gain insight into performance issues of applications running in clouds. This is achieved by a number of innovations with respect to the monitoring, modelling and managing of virtualised computing systems: (i) Monitoring – we develop a monitoring and resource control platform that, unlike early cloud benchmarking systems, enables service level objectives (SLOs) to be expressed graphically as Performance Trees; these source both live and historical data. (ii) Modelling – we develop stochastic models based on Queueing Networks and Markov chains for predicting the performance of applications in multicore virtualised computing systems. The key feature of our techniques is their ability to characterise performance bottlenecks effectively by modelling both the hypervisor and the hardware. (iii) Managing – through the integration of our benchmarking and modelling techniques with a novel interference-aware prediction model, adaptive on-line reconfiguration and resource control in virtualised environments become lightweight target-specific operations that do not require sophisticated pre-training or micro-benchmarking. The validation results show that our models are able to predict the expected scalability behaviour of CPU/network intensive applications running on virtualised multicore environments with relative errors of between 8 and 26%. We also show that our performance interference prediction model can capture a broad range of workloads efficiently, achieving an average error of 9% across different applications and setups. We implement this model in a private cloud deployment in our department, and we evaluate it using both synthetic benchmarks and real user applications. We also explore the applicability of our model to both hypervisor reconfiguration and resource scheduling. The hypervisor reconfiguration can improve network throughput by up to 30% while the interference-aware scheduler improves application performance by up to 10% compared to the default CloudStack scheduler.