Parallel Matrix Multiplication (C++ | Linux | Multiprocessing)

This project explores high-performance parallel matrix multiplication using process-level parallelism in Linux. It implements and compares different inter-process communication (IPC) strategies and workload scheduling techniques.

---

Key Ideas

• Parallel computation using fork() (multi-process model)
• Block-based (tiled) matrix multiplication
• Inter-process communication (IPC)
• Load balancing and scheduling strategies
• Performance vs communication trade-offs

---

Implementations

1. Pipe-Based IPC (Dynamic Scheduling)

• Uses pipes for communication between parent and worker processes
• Parent distributes tasks dynamically (work-stealing style)
• Workers compute matrix blocks and send results back via pipes

Highlights:

• Dynamic load balancing
• Better CPU utilization
• Higher communication overhead

---

2. Shared Memory (Static Partitioning)

• Uses System V shared memory (shmget, shmat)
• Each process writes directly to a shared result matrix
• Work is statically divided among processes

Highlights:

• Low communication overhead
• Lock-free design (disjoint memory writes)
• Simpler but less adaptive than dynamic scheduling

---

Design Decisions

Block-based Computation

• Matrix divided into smaller tiles (block_size)
• Improves cache locality and parallelism

Lock-Free Writes

• Each process writes to a unique region of the result matrix
• Avoids synchronization overhead

CPU Affinity

• Processes pinned to specific cores using sched_setaffinity
• Improves cache performance and reduces context switching

---

Correctness

• Results verified using a sequential implementation
• Floating-point comparison with tolerance (epsilon = 1e-6)

---

How to Run

g++ -o parallel_mm file_name.c -lm
./parallel_mm <matrix_size> <num_cores>