A parallel solver for Laplace's PDE. Built using OpenMP.
Compile on GCC using the flags lm, fopenmp in C99 mode.
gcc steadyStateHeat.c -lm -fopenmp -std=C99 -o steadyStateHeat
Command line arguments:
- m : Grid width
- n : Grid height
- p : Number of threads to use.
For example, if you want to solve on a 1024x1024 grid using 16 processors run the following:
./steadyStateHeat 1024 1024 16
The solver works on the Gauss Seidel Approach.
The serial algorithms is quite simple. To approach the final solution, each point in the grid takes up the value which is average of its surrounding 4 neighbors (Left, Right, Up, Down). This step is known as a relaxation. At every iteration all points in the grid are relaxed. This goes on until either we complete a specific number of iterations or if the difference between any 2 neighboring points is no more than e (epsilon), which is a preset value.
for n in range(N_ITER):
for i in range(M):
for j in range(N):
A[i][j] = average(A[i-1][j] ,A[i+1][j] ,A[i][j+1] ,A[i][j-1])
The main problem in parallelising this algorithm is the inherent data dependency. To apply relaxation on A[i][j], we need to first relax its 4 neighbors A[i-1][j] ,A[i+1][j] ,A[i][j+1] ,A[i][j-1]. Therefore we cannot do a simple row or column based data decomposition. Instead we need to do something called Red Black Decomposition. (Imagine a chessboard, red and black). All red nodes depend only on black nodes, and vice versa. There is no dependence between 2 nodes of same color. Using this data decomposition we can apply the following logic:
for n in range(N_ITER):
# iterate through all black nodes parallely
<!--BEGIN PARALLEL REGION-->
for i in range(M/2):
for j in range(N/2):
m = 2*i
n = 2*j
A[m][n] = average(A[m-1][n] ,A[m+1][n] ,A[m][n+1] ,A[m][n-1])
# iterate through all red nodes parallely
<!--BEGIN PARALLEL REGION-->
for i in range(M/2):
for j in range(N/2):
m = 2*i + 1
n = 2*j + 1
A[m][n] = average(A[m-1][n] ,A[m+1][n] ,A[m][n+1] ,A[m][n-1])
- Chahak Mehta
- Amarnath Karthi