GPU Workshop: July, 2010 Scott Briggs PhD Candidate Civil/Env. Engineering Contaminant Hydrogeology...

Post on 18-Dec-2015

219 views 1 download

Tags:

transcript

GPU Workshop: July, 2010

Scott BriggsPhD Candidate Civil/Env. Engineering

Contaminant HydrogeologySupervisors: B. E. Sleep and B. W. Karney

Contaminant Hydrogeology• Study and management of

groundwater resources.• We use computer models to

determine the best approach and expected results of a given system.

• Research specialization in zones of fractured rock using bioremediation.

• Bioremediation: the degradation of contaminants to natural or safe levels. (ex. Hydrocarbons, chlorinated solvents)

Lattice Boltzmann Methods for Modeling Rock Fractures

• Fluid flow emerges from the simulation of the intrinsic particle streaming and collision processes.

• Can incorporate micro-scale interactions:– Changing and complex boundaries.– No-slip condition.– ‘External’ forces – such as gravity and/or biofilm-fluid

interactions.

• Parallelization of LBM algorithms:– Minimal overhead due to discretized domain and locality

requirements of LBM.

3

Lattice Boltzmann Method: D2Q9

4

Sukop and Thorne, 2005

Parallel Plate Validation

Single Precision• 7.4 % relative error

Double Precision (below)• 0.78% relative error

Backward facing step Validation• Qualitative results equal to those of Armaly et al.

(1983)• Re = 100: Reattachment at 3 Step heights

• Re = 150: Reattachment at 4 Step heights

• Re = 200: Reattachment at 5 Step heights

Cubic Law in Rock Fracture Flow• The cubic law is an approximation of the N-S equations

for laminar flow through parallel plates• Traditionally the cubic law has been used in rock

fracture hydrogeology.• However there was a need to account for:– Surface roughness at varying scales– Inertial effects due to tortuosity of fracture– Contact area in 3D

• Method of comparison:• Take cubic law: • Compare flow rates between model and cubic law.

L

hWa

gQ

3)2(

12

Rock Fracture Sample #1 Flow Comparison to Cubic Law

• Flow rate: 8.1% deviation for Re of 0.06, .6 and 6.– Re = 60 deviation of 10%– Re = 600, deviation of 20% (τ approaching 0.5)– Brush and Thompson (2003) found 10% deviation

from cubic law using Stokes (low Re) simulations.

Rock Fracture Sample #2 Flow Comparison to Cubic Law

• Flow rate: 50-55% deviation for Re = 0.0006, through Re = 60.– Brown (1987) found the Cubic law to hold within

50%– Tsang (1984) suggested a order of magnitude or

more variation could occur due to tortuosity.

Rock Fracture Flow Insights• Clearly the literature is divided about the cubic

law, as are our results.• Exactly why we chose LBM and the use of the

GPU made is possible.• LBM method allows for much more insight into

the flow dynamics within the fracture, something not allowed by cubic law approximation.

• Bioremediation:

Performance Results

• Metric: Million Lattice Updates Per Second (MLUPS)• Typical CPU today: 6.2 MLUPS• Typical Single precision CUDA: 400 MLUPS (Tolke,

2008).– Single precision– Geforce 8800 Ultra

• Our code for a similar grid size: 46.2 MLUPS – Double precision– Geforce 260 Core 216

• Remember double precision = 1/8 single precision

Future Work

• Bioremediation: Implementation of bacterial populations dynamics on GPU.

• Implementation of random number generator needed for above.

• Optimization on Fermi.• Generally reduce resource requirements and

‘branchyness’ of code.

Thanks