Parallel Systems

Parallel programming coursework: eight problems in C, implemented across pthreads, OpenMP, MPI, AVX2 intrinsics and CUDA, each with benchmarks.

Directory	Problem	Models
`polynomial-multiplication/`	Dense polynomial product, O(n²) inner loop	pthreads, OpenMP, MPI, AVX2
`spmv-csr/`	Sparse matrix-vector product over a compressed sparse row structure	OpenMP, MPI
`mergesort/`	Top-down mergesort	OpenMP tasks
`synchronization/`	Four microbenchmarks: atomics vs locks, false sharing, lock granularity, barrier implementations	pthreads
`fdtd-cuda/`	2D finite-difference time-domain Maxwell solver, GPU port and optimization	CUDA, Thrust, CUB

Each directory has its own README with the implementation notes and measured results.

Build

Requires a C compiler with OpenMP support (gcc or clang). The AVX2 target needs x86-64 hardware, and the MPI targets need mpicc.

Every program has its own Makefile and builds on its own with make -C <dir>. The top level builds them in groups:

make cpu    # everything except MPI
make mpi
make        # both
make clean

The CUDA sources are notebook cells from an NVIDIA training environment and are not part of this build.

Two environments, not interchangeable:

Shared memory (pthreads, OpenMP, AVX2): AMD Ryzen 7 5700U, 16 cores, Linux 6.17.8, gcc 15.2.1.
MPI: department lab machines linux01-linux30, one process per node, so every rank boundary is a network hop.

The same problem gives opposite answers in the two environments. polynomial-multiplication/ has that comparison.