CFD Modeling of Pressure Drop through an OCP Server for Data Center Applications

Overview

The study asks how to calculate quickly and accurately the turbulent pressure drop that internal components cause in servers in air-cooled data-centre simulations. A fully resolved Open Compute Project (OCP) server was run in OpenFOAM in a virtual wind tunnel (5.98 million cells, k-omega SST) over a range of airflows, with fans stationary (inactive) and with fans included (active). The tunnel inlet was 0.5 L upstream and the outlet L downstream of the server (L = server length), with 5% inlet turbulence intensity and slip side walls.

The fully resolved snappyHexMesh model had 5.98 million cells; the porous replacement used a 20,352-cell block mesh, about 0.3% of that. Twenty simulations were run for flow rates of 0.003 to 0.06 m³/s, and steady-state results matched time-averaged transient results. Fitted Darcy-Forchheimer coefficients were: inactive server D = 782,803.905 and F = 37.395; active server D = 715,633.617 and F = 24.788. The main pressure losses occur at the inlet, at the rear heat sink and at the fans/outlet, which gives the highest drop.

A literature approach assuming porosity 0.5 and hydraulic diameter 0.55 departed strongly from the resolved pressure-drop curve, so the choice of porosity coefficients drives accuracy. The porous-versus-complete uncertainty index was largest at low flow, 0.6699 (inactive) and 0.3725 (active) at 0.003 m³/s, and fell to 0.0451 (inactive) and 0.1329 (active) at 0.06 m³/s. The authors publish both the coefficients and the OpenFOAM cases.

Limits: the coefficients are specific to one OCP server design and must be re-derived for another server geometry; uncertainty of the porous model is high at low flow, and the study treats pressure drop only, with thermal modelling out of scope.

Key points

1. The virtual wind tunnel inlet was 0.5 L upstream and the outlet L downstream of the server (L = server length), with 5% inlet turbulence intensity and slip side walls.
2. The fully resolved snappyHexMesh model had 5.98 million cells; the porous replacement used a 20,352-cell block mesh, about 0.3% of that.
3. Twenty simulations were run for flow rates of 0.003 to 0.06 m³/s, and steady-state results matched time-averaged transient results.
4. Fitted Darcy-Forchheimer coefficients: inactive server D = 782,803.905 and F = 37.395; active server D = 715,633.617 and F = 24.788.
5. A literature approach assuming porosity 0.5 and hydraulic diameter 0.55 departed strongly from the resolved pressure-drop curve, so the choice of porosity coefficients drives accuracy.
6. Porous-versus-complete uncertainty index was largest at low flow, 0.6699 (inactive) and 0.3725 (active) at 0.003 m³/s, falling to 0.0451 (inactive) and 0.1329 (active) at 0.06 m³/s.
7. The main pressure losses occur at the inlet, at the rear heat sink and at the fans/outlet, which gives the highest drop.

Using it in design

• When modelling a server as a porous block in room-scale CFD, use Darcy-Forchheimer coefficients fitted to a resolved pressure-drop curve rather than generic assumptions such as porosity 0.5.
• Use separate coefficients for servers with fans running (active) and stationary (inactive) (D = 715,633.617 and 782,803.905), and note that uncertainty is high at low flow.
• Cut server cell count to about 0.3% of the full geometry to make solving many servers in a room model practical.

Who should read it

Data-centre thermal engineers and researchers building server models for room-scale CFD.

Document contents

1. 1. Introduction
2. 2. Methods (2.1 Governing Equations, 2.2 Computational Domain, Boundary Conditions and Mesh, 2.3 Numerical Method)
3. 3. Results (validation, time-averaging, inlet conditions)
4. 4. Conclusions

Related standards

• OCP Open Rack

Read the document

Publication details

Authors

Aras Dogan; Sibel Yilmaz; Mustafa Kuzay; Cagatay Yilmaz; Ender Demirel

Publisher

Energies (MDPI AG)

Year

2022

Volume

15

Issue

17

Pages

6438

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.3390/en15176438

Licence

CC BY 4.0 — share with attribution

Licence evidence

www.mdpi.com

Citation