Reduced order thermal modeling of data centers via proper orthogonal decomposition: a review

Abstract

PurposeThe purpose of this paper is to review the available reduced order modeling approaches in the literature for predicting the flow and specially temperature fields inside data centers in terms of the involved design parameters.Design/methodology/approachThis paper begins with a motivation for flow/thermal modeling needs for designing an energy‐efficient thermal management system in data centers. Recent studies on air velocity and temperature field simulations in data centers through computational fluid dynamics/heat transfer (CFD/HT) are reviewed. Meta‐modeling and reduced order modeling are tools to generate accurate and rapid surrogate models for a complex system. These tools, with a focus on low‐dimensional models of turbulent flows are reviewed. Reduced order modeling techniques based on turbulent coherent structures identification, in particular the proper orthogonal decomposition (POD) are explained and reviewed in more details. Then, the available approaches for rapid thermal modeling of data centers are reviewed. Finally, recent studies on generating POD‐based reduced order thermal models of data centers are reviewed and representative results are presented and compared for a case study.FindingsIt is concluded that low‐dimensional models are needed in order to predict the multi‐parameter dependent thermal behavior of data centers accurately and rapidly for design and control purposes. POD‐based techniques have shown great approximation for multi‐parameter thermal modeling of data centers. It is believed that wavelet‐based techniques due to the their ability to separate between coherent and incoherent structures – something that POD cannot do – can be considered as new promising tools for reduced order thermal modeling of complex electronic systems such as data centersOriginality/valueThe paper reviews different numerical methods and provides the reader with some insight for reduced order thermal modeling of complex convective systems such as data centers.

Forced convection heat transfer within a moderately‐stenosed, patient‐specific carotid bifurcation

“…Although forced convection heat transfer is not a new topic (Mauro et al, 2010;Al-Amri and El-Shaarawi, 2010;Codina et al, 2010;Rainieri et al, 2011;Chennu and Paturu, 2011;Drikakis and Asproulis, 2010;Jaluria, 2010;Lai et al, 2011;Manca et al, 2011;Samadiani and Joshi, 2010;Sunden, 2010), application to a patient-specific carotid geometry is new in the present study. A numerical method that has been proven to be both robust and accurate is the characteristic based split (CBS) scheme (Nithiarasu, 2003;Nithiarasu et al, , 2012Bevan et al, 2012).…”

Coupled inviscid-viscous solution method for bounded domains: Application to data-center thermal management

“…In order to increase computational efficiency, reduced order modeling techniques have been used such as ad hoc methods [17][18][19][20], Proper Orthogonal Decomposition (POD) [21], and various inviscid and potential flow methods [13,14,22,23]. In general, these methods increase the computational efficiency relative to traditional CFD/HT RANS modeling methods, but at the cost of modeling accuracy.…”

“…POD based algorithms have proven quite effective in reducing the computational effort, along with providing reasonably accurate results at the rack scale, but have not been able to provide accurate results at the room scale [21]. In combination with ad hoc methods, a POD model has shown both increased computational speed and accuracy for a specific data-center layout [24].…”

Feedback control of linear distributed parameter systems via adaptive model reduction in the presence of device network communication constraints

“…We can now define the error difference between states of the process of (19) and the model of (14) as follows…”

“…POD variants have been used extensively in model reduction (see reviews [8], [13], [19] and references therein) and optimization and control of DPS [1], [2], [10], [15]. The POD approach assumes a priori availability of a sufficiently large Financial support from the National Science Foundation, CMMI Award # 13-00322 is gratefully acknowledged.…”