School of something School of Mechanical Engineering FACULTY OF OTHER ENGINEERING Computational Fluid Dynamic Investigation of Liquid Rack Cooling in Data Centres A. Almoli 1, A. Thompson 1, N. Kapur 1, J.L. Summers 1, H.M. Thompson 1 and G. R. Hannah 2 1 School of Mechanical Engineering, University of Leeds 2 Airedale Air Conditioning Ltd SusTEM Conference 2010, Newcastle, 3 rd November 2010
Outline What is a Data Centre? Energy Challenges in Data Centres CFD Modelling Methodology Comparison of Passive and Active Back Door Coolers Conclusions 3 rd November 2010 2
Data centres: What is a Data Centre? Critical to digital economy and society Produce 25% of world s total IT related emissions Year on year growth of ~16% Typically each kw of IT power used requires 1kW for cooling UK s financial centres mean UK currently internationally leading 3 rd November 2010 3
What is a Data Centre? To maintain the condition of air entering the unit (temperature and humidity) cooling is essential: Hot aisle / cold aisle air is distributed from one side of racks to another Water cooled units allow local cooling of each blade 3 rd November 2010 4
Energy Challenges in Data Centres Inefficiencies in IT and Cooling: Roughly each KWh for data processing requires 1KWh for cooling. IT inefficiencies addressed by: Improved semi-conductor technologies Server virtualisation Cooling Inefficiencies: Most popular approach: cold air and containment BUT can also use: Direct liquid cooling of servers Dielectric immersioon On-chip cooling 3 rd November 2010 5
Energy Challenges in Data Centres http://www.helen.fi/pdf/computer_hall.pdf 3 rd November 2010 6
CFD Modelling Methodology Data Centres are Critical Environments where catastrophic data losses must be avoided. These can be caused by: Over-heating servers Hygroscopic dust Electric discharge failures CFD is increasingly being used to improve air flow design in data centres. We describe a new CFD methodology for liquid loop heat exchangers back door coolers Investigate potential benefits of active (with fans) versus passive (no fans) back door coolers 3 rd November 2010 7
CFD Modelling Methodology Thermal air flows in Data Centres are complex, recirculating flows with a hierarchy of length scales: Typical Reynolds number based on: Air inlet velocity 1m/s Rack length scale 2.4m Re 100,000 BUT Re through servers much smaller - flow through servers has not been described adequately As a result, CFD is still largely unverified for Data Centre air flows 3 rd November 2010 8
CFD Modellng Methodology We consider a 3 aisle, six rack data centre considered for illustrative purposes. 3 rd November 2010 9
CFD Modelling Methodology Cold aisle represented by Region B: To left is composed of: Server rack (rack 1) Back door cooler (back door cooler 1) Hot aisle (region A) 3 rd November 2010 10
CFD Modelling Methodology Previous CFD studies have ignored small-scale features Here, treat rack as porous medium: define source term in the momentum equations in terms of permeability tensor define energy source term from manufacturers specifications Heat load per rack (kw) Source term per rack (W/m 3 ) 15 7102.2 25 11837 30 14204.5 3 rd November 2010 11
CFD Modelling Methodology Air flow through a server rack aided by server fans modelled by pressure drop across rear rack boundary.. Based on Comair Rotron fan: P = 2 948.4 131.39V + 16.2V 1.154V 3 (for 3 lower server blocks) P = 2 1481.87 164.23V + 16.2V 0.923V 3 (for upper server block) Fan speed and pressure drop expressed via: P P 1 2 N = N 1 2 2 3 rd November 2010 12
CFD Modelling Methodology Back Door Cooler modelled as a porous medium and as a radiator which removes energy from air stream at rate: Q& bdc = m& h ( ) U A T T ) ( ref air Based on experimental results of Tang et al (2009), the heat transfer coefficient, h ( ) 0. 71 U U and the constant of proportionality is 900 and encompases air properties and fin and tube geometries. It also is based on removing most of the thermal energy from the 30kW rack. 3 rd November 2010 13
Air Flow within the Data Centre Air enters through air vents at temperature T IN from CRAC units velocity vectors temperature distribution Cool air rises in cold aisle, is drawn through servers racks and is heated by IT loads and passes into hot aisles at the back of servers 3 rd November 2010 14
Comparison of Passive and Active Back Door Coolers E.g. Each of six racks adds 30kW into air streams - 180kW Without back doors CRAC units must remove full 180kW With passive back door CRAC units must remove 14.5kW With active back doors CRAC units must remove only 11.5kW 3 rd November 2010 15
Comparison of Passive and Active Back Door Coolers CFD predictions of rack intake and exhaust temperatures at various heat dissipation rates of 15 and 30kW: Exhaust air temperatures leaving active back door are lower Typically between 1 and 2 o C for 30kW case with parameters used here 3 rd November 2010 16
Conclusions Increasing energy demands of Data Centres presenting industry and governments with an energy supply problem reliance on air as heat transfer medium is problematical CFD modelling for Data Centres is largely unvalidated and treatment of server racks is crucial, although few details have appeared in the literature to date. New methodology proposed for simulating effect of liquid loop heat exchanger at the rear of server racks enables additional benefits of active back doors to be investigated Detailed validaton of the CFD methodology is needed for practical scenarios 3 rd November 2010 17