Zhihang Song

A Compact Thermal Model for Data Center Analysis using the Zonal Method

Modeling the thermal environment for data centers, including prediction of airflow and temperature distributions, is generally an extremely time-consuming process using full-scale CFD analysis. Reduced order models are necessary in order to provide real-time assessment of cooling requirements. The use of a coarse-grained zonal model is being investigated as a predictive tool, and this article details the development and implementation of a three-dimensional, pressurized zonal model. To construct and validate the zonal model, a basic data center configuration was analyzed using a finite volume software package. The calculated flow fields provide the spatial flow coefficients required in the zonal model, which is based on the power law method (PLM). A physically-based mapping between the controllable spatial mass flow rate and temperature distribution was obtained. Good agreement (within 10% average relative error) was obtained between the zonal model predictions and the CFD results. These preliminary results show promise that zonal models may yield an effective real-time thermal management design tool for data centers.

Multivariate Prediction of Airflow and Temperature Distributions Using Artificial Neural Networks

Thermal Management optimization for data centers, including prediction of airflow and temperature distributions, is generally an extremely time-consuming process using full-scale CFD analysis. Reduced order models are necessary in order to provide real-time assessment of cooling requirements for data centers. The use of a simulation-based Artificial Neural Network (ANN) is being investigated as a predictive tool. A model for a basic hot aisle/cold aisle data center configuration was built and analyzed using the commercial software FloTHERM. The flow field and temperature distributions were obtained for 100 representative sets of operating conditions using the CFD package. The Latin Hypercube Sampling technique was employed to select values for three design variables: plenum height, percentage open area of the perforated tiles and air leakage fraction. The FloTHERM results were used to generate a database for the ANN training. The CFD results from 85 cases were used for training and 16 cases were used for validation. A multivariate mapping between the input design variables and output variables (individual tile flow rates and maximum rack temperatures) was obtained. Good agreement (0.5% average relative error) was obtained between the ANN model predictions and the CFD results. These preliminary results are promising and show that an ANN based model may yield an effective real-time thermal management design tool for data centers.Copyright

Multi-objective optimization of temperature distributions using Artificial Neural Networks

Modeling the thermal environment of data centers, including prediction of the air flow and temperature distributions can be computationally intensive using CFD. Reduced order models or data-driven meta-models are necessary to provide real-time assessment of optimum operating conditions for data centers to reduce energy usage. Here, a simulation-based Artificial Neural Network (ANN) approach is employed as a predictive tool. A model for a basic single cold aisle data center configuration is analyzed using the commercial CFD software FloTHERM. The simulation results are used to generate a database for training and cross validation of a primary ANN corresponding to a specific set of input and output operating conditions. Good agreement is achieved between the CFD and ANN based model predictions for maximum rack inlet temperatures over a range of operating conditions. In addition, by combining the ANN with a cost function based Multi-Objective Genetic Algorithm (MOGA), the operating conditions can be inversely predicted for desired outputs (e.g. rack inlet temperatures). The total simulation time for the ANN-MOGA approach is reduced significantly compared to a fully CFD-based optimization methodology.

Parametric analysis for thermal characterization of leakage flow in data centers

A model for a basic hot aisle/cold aisle data center configuration was built and analyzed using the commercial computational fluid dynamics (CFD) software FloTHERM. The CFD study was motivated by experimental data that sho wed air recirculatio n from the hot aisle to the cold aisle through the gap between the floor and the bottom of the server cabinet. This flow can be attributed to lower values of the local pressure at the front of the cabinet due to the air stream from the raised-floor perforated tiles. Here this “leakage” flow is modeled to gain better understanding of the characteristics of the undesirable hot fluid recirculation. The parameters and conditions varied in the numerical investigation include: tile perforation area, CRAC provisioning, pressure gradient of the leakage air stream and the effect of cold aisle containment. The impact of the leakage flow on the cooling air supplied to the servers is discussed. The results indicate that the cooling effectiveness in data centers can be reduced due to even a small amount of under-cabinet leakage, although better cooling performance is obtained using a contained cold aisle. The leakage effect is found to be sensitive to under-provisioned conditions. The results of the numerical study can be used to better optimize the cooling performance for contained or uncontained cold aisles when there is under cabinet leakage flow that influences the server inlet air temperature.

Prediction of Hot Aisle Partition Airflow Boundary Conditions

The integration of a simulation-based Artificial Neural Network (ANN) with a Genetic Algorithm (GA) has been explored as a real-time design tool for data center thermal management. The computation time for the ANN-GA approach is significantly smaller compared to a fully CFD-based optimization methodology for predicting data center operating conditions. However, difficulties remain when applying the ANN model for predicting operating conditions for configurations outside of the geometry used for the training set. One potential remedy is to partition the room layout into a finite number of characteristic zones, for which the ANN-GA model readily applies. Here, a multiple hot aisle/cold aisle data center configuration was analyzed using the commercial software FloTHERM. The CFD results are used to characterize the flow rates at the inter-zonal partitions. Based on specific reduced subsets of desired treatment quantities from the CFD results, such as CRAC and server rack air flow rates, the approach was applied for two different CRAC configurations and various levels of CRAC and server rack flow rates. Utilizing the compact inter-zonal boundary conditions, good agreement for the airflow and temperature distributions is achieved between predictions from the CFD computations for the entire room configuration and the reduced order zone-level model for different operating conditions and room layouts.Copyright

Improved Zonal Model for Data Center Analysis

A newly constructed zonal model based on the velocity propagation method (VPM) was developed as a thermal analysis tool for data centers. A viscous loss model is included, to better account for airflow momentum instead of using the basic power law method (PLM). The zonal model is implemented in the equation-based and object-oriented environment SPARK. A CFD model for a single cold aisle server room configuration was built and analyzed using the commercial software FloTHERM. Cold aisle containment was studied. Results from the zonal model, including both the plenum flow field and rack inlet temperature distributions were compared with those from the CFD package. Good agreement (within 10% average relative error) was obtained between the zonal model predictions and the CFD results. A primary goal of the study is to develop an effective real-time thermal analysis tool based on the zonal approach.Copyright

Data Center Transient Flow Analysis Using Proper Orthogonal Decomposition

Real-time analysis of the transient temperature distribution and flow field in a data center is not possible using well-resolved computational fluid dynamics (CFD) simulations. Reduced order models must be used to predict the optimum operating and control conditions to achieve better energy-efficiency. Here, the proper orthogonal decomposition (POD) method is used to model transient behavior of a simple hot aisle/cold aisle data center configuration. Verified CFD simulation results were used to generate snapshots for building a reduced order POD model corresponding to transient variation of the computer room air conditioner (CRAC) operating conditions. Good agreement is achieved between the CFD and reduced order model predictions for the evolving flow structure over a range of CRAC supply operating conditions. Once constructed, the computational time required to obtain the POD results for the transient response is considerably reduced compared to the CFD simulations. The advantages and disadvantages of the POD method for this type of transient behavior are discussed, and recommendations are made on using this type of compact modeling approach to develop a real-time predictive tool for data center design and control to enhance energy efficiency.Copyright