A STATISTICAL APPROACH TO REAL-TIME UPDATING AND AUTOMATIC SCHEDULING OF PHYSICAL MODELS Huijing Jiang ∗ Research Staff Member Business Analytics & Mathematical Sciences IBM Thomas J. Watson Research Center Yorktown Heights, NY 10598 Email: huijiang@us.ibm.com Xinwei Deng Assistant Professor Department of Statistics Virginia Tech Blacksburg, VA 24061 Email: xdeng@vt.edu Vanessa Lopez Research Staff Member Business Analytics & Mathematical Sciences IBM Thomas J. Watson Research Center Yorktown Heights, NY 10598 Email: lopezva@us.ibm.com Hendrik Hamann Manager Photonics and Thermal Physics IBM Thomas J. Watson Research Center Yorktown Heights, NY 10598 Email: hendrikh@us.ibm.com ABSTRACT Energy consumption of data center has increased dramati- cally due to the massive computing demands driven from every sector of the economy. Hence, data center energy management has become very important for operating data centers within en- vironmental standards while achieving low energy cost. In order to advance the understanding of thermal management in data centers, relevant environmental information such as temperature, humidity and air quality are gathered through a network of real- time sensors or simulated via sophisticated physical models (e.g. computational fluid dynamics models). However, sensor read- ings of environmental parameters are collected only at sparse locations and thus cannot provide a detailed map of temperature distribution for the entire data center. While the physics mod- els yield high resolution temperature maps, it is often not fea- sible, due to computational complexity of these models, to run them in real-time, which is ideally required for optimum data ∗ Address all correspondence to this author. center operation and management. In this work, we propose a novel statistical modeling approach to updating physical model outputs in real-time and providing automatic scheduling for re- computing physical model outputs. The proposed method dy- namically corrects the discrepancy between a steady-state out- put of the physical model and real-time thermal sensor data. We show that the proposed method can provide valuable informa- tion for data center energy management such as real-time high- resolution thermal maps. Moreover, it can efficiently detect sys- tematic changes in a data center thermal environment, and auto- matically schedule physical models to be re-executed whenever significant changes are detected. NOMENCLATURE N Total number of grid-points in physical models. n Total number of sensor locations. T Total number of time points. x M (s) Physical model output at location s. Proceedings of the ASME 2013 International Technical Conference and Exhibition on Packaging and Integration of Electronic and Photonic Microsystems InterPACK2013 July 16-18, 2013, Burlingame, CA, USA IPACK2013-73042 1 Copyright © 2013 by ASME Downloaded From: http://heattransfer.asmedigitalcollection.asme.org/ on 01/05/2015 Terms of Use: http://asme.org/terms