The Evaluation Of Servers’ Performance Degradation For Power Budget Optimization

With the development of Internet, Internet Service Providers(ISP) purchase thousands or millions servers to provide their services. In order to get rid of maintaining cost, it is popular for ISPs to deploy their servers in leased datacenter from network operators. However, in the datacenter leased to ISP, low rack utilization has seriously impaired the benefit of ISP, which incurs great lost in leasing cost. With the investigation in real datacenter, the over-provisioning in power budget is the main reason causing low rack utilization.To solve the problem that power budget optimization may bring about unpredictable applications’ performance degradation, this research deeply explore the form of performance degradation by scenario analysis and formula derivation. Based on the CPU utilization trace of servers, this is a fine-grained evaluation method for applications’ performance, which contains two metrics, describing in what percentage the application is affected by the delay and how long the delay is, respectively. In order to calculate the degree of performance degradation efficiently, the evaluation method fulfills an algorithm working iteratively on the CPU utilization trace, which can obtain the performance degradation of thousands of servers in datacenter in a very low time and space complexity. By the use case in datacenters, an approach of how this evaluation method should be used is proposed. Administrators provide the CPU utilization trace and utilization-power relationship of servers to the module integrating the evaluation method, then the module will output the performance degradation under different power budgets, and will yield a power budget allocation list instructing the power budget optimization based on users’ acceptable performance degradation.The test is conducted on the simulation of real traces in datacenters. Results show that the performance degradation evaluation used in the power budget optimization achieves more than 90% accuracy in evaluation. Meanwhile, if applying the optimization method in practical datacenters, there is at least 1/3 revenue in improving the inefficient utilization of power provisioning todays, while guaranteeing minor effects on applications.