Managing energy, performance and cost in large scale heterogeneous datacenters using migrations
Future generations computer systems
Improving datacenter energy efficiency becomes increasingly important due to energy supply problems, fuel costs and global warming. Virtualisation can help to improve datacenter energy efficiency through server consolidation which involves migrations that can be expensive in terms of extra energy consumption and performance loss. This is because, in clouds, Virtual Machines (VMs) of the same instance class running on different hosts may perform quite differently due to resource heterogeneity.
... a result of variations in performance, different runtimes will exist for a given workload, with longer runtimes potentially leading to higher energy consumption. For a large datacenter, this would both reduce the overall throughput, and increase overall energy consumption and costs. In this paper, we demonstrate how the performance of workloads across different CPU models leads to variability in energy efficiencies, and therefore costs. We investigate through a number of experiments, using the Google workload traces for 12,583 hosts and 492,309 tasks, the impact of migration decisions on energy efficiency when performance variations of workloads are taken into account. We discuss several findings, including (i) the existence of a trade-off between overall energy consumption and performance (hence cost), (ii) that higher utilization decreases the energy efficiency as it offers fewer chances to CPU management tools for energy savings, and (iii) how our migration approach could save up to 3.66% energy, and could improve VMs performance up to 1.87% compared with no migration. Similarly, compared with migrate all, the proposed migration approach could save up to 2.69% energy, and improve VMs performance up to 1.01%. We discuss these results for different combinations of VM allocation, migration policies and different benchmark workloads 1 . model resource and workload heterogeneities in the context of a cloud platform [Sec. 4]; 2. an approach/metric to balance/measure the trade-off involved in energy consumption and performance (hence cost); 3. an energy-performance-cost (Epc-aware) consolidation approach [Sec. 5.1]; 4. large-scale simulations using a real dataset from a cloud provider -Google [Sec. 6]; and 5. a review of state-of-the-art energy-performance-cost efficient scheduling techniques in infrastructure clouds [Sec. 2]. The rest of the paper is organized as follows. We offer an overview of the related work in Sec. 2. In Sec. 3, we discuss the VM allocation process as a multi-objective optimization problem 3