Poster 108: Power Prediction for High-Performance Computing
TimeThursday, 21 November 20198:30am - 5pm
DescriptionExascale computers consume large amounts of power both for computing and cooling-units. As power of the computer varies dynamically corresponding to the load change, cooling-units are desirable to follow it for effective energy management. Because of time lags in cooling-unit operations, advance control is inevitable and an accurate prediction is a key for it. Conventional prediction methods make use of the similarity between job information while in queue. The prediction fails if there is no previously similar job. We developed two models to correct the prediction after queued jobs start running. By taking power histories into account, power-correlated topic model reselects more suitable candidate and recurrent-neural-network model considering variable network sizes predicts power variation from shape features of it. We integrated these into a single algorithm and demonstrated high-precision prediction with an average relative error of 5.7% in K computer as compared to the 18.0% obtained using the conventional method.