Free Surfaces with LBM D3Q19

Project for Computational Fluid Dynamics Lab course.

Group 6: Nathan Brei, Irving Cabrera, Natalia Saiapova.

Compilation.

Requirements: UNIX based OS.

Use make command to compile:

>> make

There is debug mode available. This mode includes additional checks for velocity and density, computes total mass in the domain and checks that there is no FLUID cell which is neighboring with a GAS cell directly. Also, it calculates the time spent on the following parts: streaming, collision, flag update and boundary treatment.

>> make debug

Usage

To run the program use

>> ./lbsim <exampleName> <number of threads>

For each example it is required to have <exampleName>.dat and <exampleName>.pgm files. They have to be placed in the examples directory. <number of threads> specifies how many threads will be used by OpenMP.

The vtk-output directory contains vtk output files. If thare is no such directory it will be created.

Description of input files

We added a few new fields to the input.dat file:

Parameter	Description
radius	Radius of a droplet
exchange_factor	Factor to increase speed of a mass exchnge (mainly to get more beautiful videos)
forces_x	Gravity force in the `x` direction
forces_y	Gravity force in the `y` direction
forces_z	Gravity force in the `z` direction

What we have done

New cell types

First of all, we added 2 new cell types: FLUID and INTERFACE. Here is table representing cell types, which should be used in <exampleName>.pgm file:

Cell type	Value	Cell type	Value
FLUID	0	PRESSURE_IN	5
INTERFACE	1	OBSTACLE	6
MOVING_WALL	2	FREESLIP	7
INFLOW	3	NOSLIP	8
OUTFLOW	4	GAS	9

Test functions

In the checks.c one can find three check functions:

check_in_rank checks that every FLUID cell has a density between 0.9 and 1.1 and norm of the velocity vector is less then sqrt(3) * C_S;
check_flags checks that no FLUID cell has a GAS neighbor;
check_mass compute total mass in the domain (FLUID and INTERFACE cells).

This checks will be activated in debug mode and be executed every time when we write output file.

External forces influence

In collision step additional term for external forces was added. Related function in collision.c: computeExternal(int i, float density, float * extForces), where extForces is a force vector specified in input file.

Mass and fluid fraction fields

For every cell we store 2 additional float values: mass and fluid fraction. We update mass field during streaming step. Fluid fraction field is updated during collide step. To be able get more fast we introduced exchange_factor in an input file. By this factor we increase mass exchange between cells.

DF from gas cells

In the streaming step was added reconstruction from GAS cells. Streaming works normally for all FLUID cells amd for INTERFACE cells we are checking whether a neighbor is GAS. If yes, we are using outflow boundary formula in this direction.

Flag update

From the mass fraction for each cell we determine whether that cell emptied or filled. We track these cells using two arrays, emptiedCells and filledCells. We perform the filling and emptying in separate phases, performFill() and performEmpty(). These phases are asymmetric in order to handle the case of a cell filling adjacent to a cell emptying. If this happens, the 'filling' operation cancels the neighbor's 'emptying' operation and the neighboring cell is stricken from emptiedCells.

Both phases maintain the loop invariant of a contiguous interface layer. When an INTERFACE cell is converted to FLUID, all GAS neighbors are converted to INTERFACE. Correspondingly, when an INTERFACE is converted to GAS, all FLUID cells are converted to INTERFACE.

Parallelization with OpenMP

We have added pragma omp parallel for to streaming, collision, treatment boundaries and update flags.

Other optimizations

At the beginning our main bottleneck was computeFeq function which took approximately 42% percent of the whole time. We splitted the Q-loop in this function in a way that compiler was able to vectorize it, then we got rid of all divisions and replaced them by corresponding multiplications. Then we changed types of all double fields to float. It gave us aproximately 2x speedup.

nasay/cfdlab