High-Fidelity Simulation of a Computer Room -.0. · High-Fidelity Simulation of a Computer Room-.0....
Transcript of High-Fidelity Simulation of a Computer Room -.0. · High-Fidelity Simulation of a Computer Room-.0....
r
High-Fidelity Simulation of aComputer Room
-.0.By Shishir'Pandya
NASA Ames Research Centerfor Hewlett-Packard Labs
Collaborators:Jasim Ahmad, EL.ORET Inc.
William Chan, NASA Ames Research CenterNeal Chaderjian, NASAAmes Research Center
I
%VOutline,
• Introduction to Columbia
• Motivation
9 Methodology
• Results
• Concluding remarks
https://ntrs.nasa.gov/search.jsp?R=20090032250 2018-07-14T15:23:25+00:00Z
Columbia
FI
F,
zs*^tias^i.;
• 10240 CPUs
20 Sgq ACtix superdusters (512 CPUS/duster)
• 2nteC gtanium(b 2 processors (1.S G?-fz)
• 1TB n1nOry/cC14tcr, 44oTB storage-
- Sgq s "N UMAffex architecture (7f finiband )
Top performance to date: 51.E TFlops
inpack benchmark
#2 on the TOPSoo Cist-
wilt in < months full -- operational5 ^ Y P
wwl nas. nasa.gov/About/Projects/Columbia/columbia.htm I
3
Advanced Supercomputing Division
4
1• r
Motivation
• Need to know if cooling is adequate• 7dentify T6g-temperature regions
• Effect of ca6Ce trays in 6or isles
Find dead-zonesF • Areas of stagriatedair
• Short-cycling• CoCd air from f loor tiles returning directCy to coolers
I' • '16ot air from racks returning to racks
' Deadline: 6 weeks
,
LAM
M-ethodology`ter
• Overset mesh approach
• Geometric model1VOfi-Vr^ii j 5Cryt- 6asedmodern,
• `Many structured meshes make up the domain of interest'
• C61'mera Grid i-oA (cGj-)
• PDE solution model• Compressi6Ce ?favier-Stakes equations
• 'fiouudary conditions
• Ov40w2 so6ver
6
Body-fitted grids• COMYCex geometry
f Geometry• %u driCateraCcefrs
Overlapping POW(
Distri, Jb-u
Units (P
17
JT Overset Mesh Approach
Simulation on an overset mesh
i. Generate surface meshes
2. Generate volume meshes
3. Specify initial conditions
4. Apply boundary conditionsf
S. Solve on each volume mesh
6. Interpolate flow data in overlapping regions
7. Repeat 4-6 until converged (steady results)
8
i3 Coolers
1r Cable Trays
Perforated FloorTiles
Mass Storage
Geometric Model
9
PDE Solution Modeln
® Overflow2 (NASA/Army developed)
102 structured overset meshes, 12 miMon points
• ,Soruti6n. of 9-,f r-Stokes! equations
• Compressi6le, viscous f rows
• fow-speed, pre-conditioner for accuracy
it Obtain steady-state
• kesult: f row field/temperature fieldG^t - 01,, io®R I
` r
4 '^ t
QApproach
• Develop geometric model
• 2D sub-floor model
• Develop necessary boundary conditions
• 3D main-floor model_
• Flow visualization
ZD sub-floorCooler exit Pipes Perforated tiles
• Simulation indicates that pies 6Cocking sub floor f l-ow haves
CittCe eect on t<e per orated tfk f row rags
• _Qterature: A fuLf three d^itn-enswnaCsu6floorsimuCation wAy pes and
conduits fount(littCe ITCM, tare variation in Poor ties except near coolers.
Karki, K.C., Radmehr, A., and Patankar, S.V., "Use of Computational Fluid Dynamics forCalculating Flow Rates Through3Perforated Tiles in Raised-Floor Data Centers,"HVAC&R Research Journal, Vol: 9, No. 2, April 2003, pp. 153-16
• REASONABLE ASSUMPTION: THE COOLER FLOW RATE
IS EVENLY DISTRIBUTED AMONG ALL THE
PERFORATED FLOOR TILES
t ^
Boundary Conditions
• Perforated tiles: Temperature, speed known
• Rack intake: Speed based on fan rating
Rack exit: Known temperatur6 increase
• Cooler intake: Known speed based on CFMrating
13
Cool-Air Floor Tile
Disk Rack
Temp of r ^: _^86.3
782
62.1
70:2 Hot Isle
11 54.0 Cold Isle
Surface T(mP i-iern re
I4
Temperature/velocity vectors1'
s_
r ^^^ 1
,J^ is
SA`•'
CPURa
PD-U
Disk Rack P^
It24r1
Temperature/velocitv vectors31 v
16
CPURack
PDU
Cab1P Tr^^r
Flow Down DraftBlockage (Streamline Short Cycling)
'^^' ,^^^'dsL^3t`^E9TF"'€'..'..k:^s'^c^5_"° vs`:•&Z?'F_^.T^^iP^#r^7a'a'?ice€-. .... > :(^.'^`t" _ ^,
is
^e
)w,tion
ers
-0
' Slight Short Cycling Temp °F
At 30,000 CFM/ISU1 78.
78.22
70.262.154.0
Streamline Short Cycling
4 ^^
# « r : `Ericls Tendfloi93 Warmyyyy A;`
84.df!": 3 p } Short cycling
fi
1
19
,._,_
drftell
a^1Sul More Cool Air Less Cool Air` Enters CPU Fan Inlet Enters CPU Fan Inlet
UM
Cooler With Cooler Without
Diverter Diverter
Possible. Modification:Dver ter on Cooler
20
n I •
'
a
^r
Concluding Remarks
°-! Used overset mesh technology to evaluate thefluid/thermal character of a computer room
- • Computed average temperature in room tocertify the adequacy of the coolers
Identified high' temperature regions and deadzones
Improved understand i ng of effect of the cable
,k trays on local temperature
Identified short-cycling
3-• ^estecfyossi6Ce inodifi"cation to reduce short-cyc6incL,
Y _