• cloud computing
• datacenters
• energy efficiency
• resource overcommitment
• scheduling
• virtual machine

Cloud Computing is one of the most growing technologies in information technology, allowing users and enterprises to reduce costs by leveraging on pay-for-use services. Infrastructural costs are only handled by Cloud Service Providers (CPS), while clients have access to compute resources on demand, leveraging on an apparently unlimited pool of resources. One of the simplest but more flexible services available via CSP is virtual machine (VM) rental, allowing users to choose among multiple available configurations (Operating Systems, GPU availability, storage performance, etc.) for their VMs. Despite the ever-growing request for cloud services, CSPs generally observe low (<30%) utilization of their computation resources, leading to resource and energy wastage, a hot topic due today attention for carbon emissions and market competitiveness.

Carbon emissions related to IT energy requirements represents a significant part of the global emissions and the ever-increasing attention to the global warming phenomenon is leading enterprises – not rarely influenced by governments or international organizations’ guidelines – to study ways to reduce their energy use or, at least, to increase the efficiency of their infrastructure and the percentage of renewable energy sources they rely on. However, the estimated trends in the growth of global datacenters consumptions makes global impacts on greenhouse gas emission a challenging task to be fulfilled.

With respect to the server utilization measurements, low average resource utilization leads to overcommitment and VM consolidation solutions, but inattentive consolidation can cause Service Level Agreement (SLA) violations if not counterbalanced by wise virtual machine migrations. This research aims to provide to CSPs a method to reduce costs by scheduling VMs over a cluster of physical machines (PMs) leveraging on VM migrations guided by VM usage monitoring and predictions to increase PM utilization without suffering from performance degradation thanks to a robust probabilistic hotspot avoidance mechanism.

The central innovation of the proposed solution is the possibility to take into consideration, during the scheduling process, multiple compatibility constraints between virtual and physical machines, such as hypervisor characteristics or special hardware requirements, that limit the number of virtual machines placeable on a given physical machine. The classic bin packing problem is then modified to consider multiple versions of the same bin, potentially characterized by different costs.

In order to assert the strength of the proposed solution, multiple simulations have been run using virtual machine traces from a real enterprise scenario and the results have been compared with the performances of the currently applied methodology. Real VM utilization data have been deeply analyzed to look for insights on typical usage levels and patterns and obtained knowledge has been used to tune algorithm parameters and improvements. Indeed, the initial work was only focused on CPU overcommitment while successive analysis showed there was significant space for RAM overcommitment, in particular for VMs with huge RAM availability whose average utilization was normally lower than 50%.

No special attention to resource usage prediction methods has been given in this work despite being useful in a real scenario. Nevertheless, real traces analysis seems to suggest that a simple “steady predictor” - i.e., assuming that VM resource requirements in the next time slice will be equals to the last observed values – could be adequate in multiple cases if paired with sufficient “safety margins”.