Download Now Free registration required
This paper takes up statistical profiling-based techniques with the aim to derive effective power provisioning in data centers. The capacity planning practices that are currently in use are based on heavy over-provisioning of power infrastructure. This shows negative impact on the operational costs of data centers as well as the computational work they can support. The paper investigates a combination of statistical multiplexing techniques to achieve better utilization of the power hierarchy within a data center. The method presented herein employs controlled under-provisioning and over-booking of power needs of hosted workloads at the highest level of the power hierarchy. While at the lower levels, the paper introduces the novel notion of soft fuses. This helps to flexibly distribute provisioned power among hosted workloads based on their needs. This combination of statistical multiplexing techniques presented by the document is based on the measurement-driven profiling and prediction framework. This helps to characterize key statistical properties of the power needs of hosted workloads and their aggregates. The improvement in the better utilization of the power hierarchy within a data center is calculated in terms of the amount of computational work (CPU cycles) per provisioned unit of power. The results show that this technique is able to double the CPW offered by a Power Distribution Unit (PDU) running the e-commerce benchmark TPC-W compared to conventional provisioning practices.
- Format: PDF
- Size: 258.3 KB