Parallel 3D Grid Processing with MPI

An MPI-based implementation for distributed processing of large 3D grid datasets using domain decomposition and ghost-cell communication. The project focuses on parallel detection of local and global extrema while exploring performance and scalability on distributed-memory systems.

Developed as part of CS633: Parallel Computing at IIT Kanpur.

---

Features

• 3D domain decomposition across MPI processes
• Ghost-cell (halo) exchange for boundary communication
• Non-blocking communication using MPI_Isend / MPI_Irecv
• MPI derived datatypes for efficient data transfers
• Support for uneven domain partitioning
• Strong scaling experiments and performance benchmarking
• Multiple implementations (baseline, optimized, and final)

---

Repository Structure

.
├── Assignment.pdf
├── LICENSE
├── README.md
├── baseline.c
├── optimised.c
├── final.c
├── src_c.c
├── test.c
├── datasets/
│   ├── data_64_64_64_3.bin
│   ├── data_64_64_64_3.txt
│   ├── data_64_64_96_7.bin
│   └── data_64_64_96_7.txt
└── results/
    ├── scaling_results.csv
    ├── data_64_64_64_3_scaling_boxplot.png
    ├── data_64_64_96_7_scaling_boxplot.png
    ├── strong_scaling_boxplot.png
    └── strong_scaling_plot.png

---

Parallelization Strategy

The implementation distributes a 3D grid across a Cartesian process grid (PX × PY × PZ).

Each MPI process:

• Owns a local subdomain of the global grid
• Exchanges ghost cells with neighboring processes
• Computes local minima and maxima
• Participates in global reductions to determine overall extrema

The implementation uses:

• Cartesian domain decomposition
• Non-blocking point-to-point communication (MPI_Isend / MPI_Irecv)
• MPI derived datatypes for efficient boundary exchange
• Support for uneven grid partitioning across processes

---

Building

Requirements

• MPI (OpenMPI or MPICH)
• GCC (C99 compatible)
• Linux

Compile

mpicc -O3 -o final.x final.c -lm

---

Running

mpirun -np <processes> ./final.x \
    <dataset> \
    <PX> <PY> <PZ> \
    <NX> <NY> <NZ> \
    <TIMESTEPS> \
    <output>

Example

mpirun -np 8 ./final.x \
datasets/data_64_64_64_3.txt \
2 2 2 \
64 64 64 \
3 \
output.txt

---

Output

The program reports:

• Local extrema counts
• Global minimum and maximum values
• Read time
• Computation time
• Total runtime

Example:

(125,143), (98,156), (110,132)
(0.125,9.875), (0.250,9.750), (0.313,9.625)
0.125,0.875,1.000

---

Performance

The project includes strong scaling experiments across multiple MPI process counts.

Key observations:

• 3D decomposition improves load balancing compared to lower-dimensional partitioning.
• Communication overhead dominates for smaller problem sizes.
• Larger datasets demonstrate improved scalability.
• Ghost-cell exchange using MPI derived datatypes minimizes packing overhead.

Performance plots and scaling results are available in the results/ directory.

---

Technologies

• C
• MPI
• OpenMPI / MPICH
• Linux
• Distributed-memory parallel computing

---

License

Released under the MIT License. See the LICENSE file for details.