Ph.D. Proposal Presentation by Jeffrey Rambo
Friday, June 11, 2004

( Dr. Yogendra Joshi, Chair)

"Model Reduction of Multiscale Turbulent Convection: Application to Data Center Thermal Management "

Abstract

Data centers are large infrastructure facilities that house data processing equipment, arranged in 2-meter tall enclosures called racks. The rapid increase in the volumetric heat generation of today’s electronics has lead to rack power dissipation levels as high as 20 kilowatts, and coupled with facilities growing as large as thousands of square meters, data centers are capable of dissipating several megawatts of power. These facilities are intended to operate continuously and provide a safe and reliable operating environment for the equipment contained within. However, only recently are these systems level thermal management issues being addressed.

The thermal management of electronics has traditionally focused on a single chip or enclosure, which is able to reject heat to an extensive ambient. In a data center, multiple enclosures reject heat to a common ambient and the interaction among the various data processing equipment can greatly influence the thermal performance. Due to complex nature of the turbulent data center airflow, computational fluid dynamics and heat transfer (CFD/HT) are required to investigate the thermal performance. The relevant heat transfer processes act over a wide range of length scales from the heat generated by the individual chips in each server to the facility level cooling scheme, making the problem inherently multiscale and resulting in large, computationally expensive CFD/HT models.

The large model size of CFD/HT simulations makes robust design studies and optimization impossible. The proper orthogonal decomposition (POD) has successfully produced low order models in a dynamical system approaches to turbulent flows and has demonstrated the capability to be adapted to steady state observations of turbulent flows. The POD extracts the fundamental modes of a complex system and currently subspace projection algorithms are being developed to produce accurate estimations of the turbulent flow fields. The reduced-order modeling framework can then be extended to estimate data center airflow and heat transfer characteristics and a generalized design methodology can ultimately be produced. A unique experimental data center facility is under construction to perform the model validation and evaluate next generation data center cooling schemes.