A predictive approach for enhancing resource utilization in PaaS clouds

Data centers have been criticized for being heavy consumers of energy, accounting for a significant share of the world's total energy use. This is compounded by the fact that a large portion of the computational resources in data centers remains idle most of the time, while nevertheless consuming a great amount of energy even in an idle state. One of the causes of resource idleness is the static resource allocation strategy of the widespread quota management systems in private clouds, which commonly do not keep track of the utilization rate of allocated resources, which are often idle. Better resource management strategies are therefore called for to improve the utilization rate of available capabilities and thereby minimize operational costs. We propose here a method for resource provisioning based on utilization-rate forecasting for clouds with data-processing PaaS environments. A clustering algorithm is used for estimating and dynamically scheduling the workload of batch applications. Based on current and historical records of resource utilization, the algorithm enables taking efficient scheduling decisions. Due to the non-proportionality of computing power consumption, we show that compared to static quota management systems for private clouds, average increases of 10+ in CPU and RAM utilization and 20+ in the number of processed jobs may be obtained without a significant reduction in the quality of service.