Description

Data center power consumption increases every year. To reduce this, it is necessary not only for individual equipment to save energy, but also to optimize the operational conditions of the equipment in the data centers such as air conditioners. Using machine learning to predict power usage in a whole data center is effective to determine optimum operational parameters. However, although accuracy is higher, this method requires building a learning model for the actual operations of the entire data center. Acquisition of training data from real operations takes time to cover all operating conditions. It is also possible to improve accuracy by using reinforcement learning, but this also requires time to gather the learning data for sufficient accuracy. Furthermore, in the case of a new data center, no learning data exists at all, so it is difficult to operate under optimal conditions in the initial stages. In addition, there is an operational problem: if servers are owned by a customer, such as those for collocation services, workload information for those servers will not be shared with data center administrators. For this reason, learning the data necessary for energy saving is not handled as an optimal operational parameter. From that viewpoint, there is a great need for power predictions that do not require prior learning of the entire data center. To predict the power consumption of the whole data center without prior learning, we proposed a power consumption simulator based on simple metrics combining a power consumption model of individual devices and computational fluid dynamics (CFD) or a simulation of air flow velocity distribution. As a result, we achieved an accuracy with only an 8% power consumption prediction error for the whole data center, in which 220 servers were implemented on seven racks. If the power consumption models of individual equipment are obtained in advance, the overall power prediction and control of the data center is possible. This method also exhibits promising potential to be a data center optimizer from the perspective of power consumption.

Citation: 2019 Winter Conference, Atlanta, GA, Conference Papers

Product Details

Published:

2019

Number of Pages:

8

Units of Measure:

Dual

File Size:

1 file , 2.7 MB

Product Code(s):

D-AT-19-C014