부산대학교 도서관

Cloud computing is leading to transformational changes with stringent requirements on usability, performance and security over very heterogeneous workloads. Their run-time management requires realistic algorithms and techniques for sampling, measurement and characterization for load prediction. Due to the expectation of elasticity, large swings in their demand are common, which cannot be modeled accurately based on raw measures such as the number of session requests, which show very large variability and poor auto-correlation. We demonstrate the use of load prediction algorithms for cloud platforms, using a two-step approach of load trend tracking followed by load prediction, using cubic spline Interpolation, and hotspot detection algorithm for sudden spikes. Such algorithms integrated into the autonomic management framework of a cloud platform can be used to ensure that the SaaS sessions, virtual desktops or VM pools are autonomically provisioned on demand, in an elastic manner. Results indicate that the algorithms are able to match representative SaaS load trends accurately. This approach is suitable to support different load decision systems on cloud platforms with highly variable trends in demand, and is characterized by a moderate computational complexity compatible to run-time decisions.