High-Performance Parallel Algorithms Implementation

Professional portfolio of parallel computing work focused on performance optimization, memory hierarchy awareness, and scalable execution across CPU, GPU, and distributed environments.

Project Portfolio

Project	Key Technique	Tech Stack	Description
hw1_odd-even-sort	Odd-even transposition	MPI, C++	Distributed odd-even sorting with message-passing coordination.
hw2_mandelbrot-set	Work partitioning	Pthreads, MPI, OpenMP, C++	Parallel Mandelbrot generation with threading and hybrid MPI/OpenMP variants.
hw3	Blocked APSP	CUDA, OpenMP, C++	Blocked Floyd-Warshall with CPU and GPU implementations.
hw4	GPU kernel optimization	CUDA, C++	CUDA-based parallel implementation for assignment-specific workload.
hw5	Communication tuning	UCX	Configuration and diff for high-performance communication setup.
lab3	Image processing	CUDA, OpenACC, C++	Sobel filter and MNIST processing across GPU and accelerator models.
lab4	Shared-memory tiling	CUDA, C++	Optimized Sobel filter using shared memory and tile-based access.

Performance Optimization Deep Dive

lab4 (CUDA Sobel Filter)

The CUDA Sobel filter is fast because it stages input tiles in shared memory, reducing repeated global memory loads for overlapping convolution windows. This increases data reuse, minimizes global memory traffic, and raises computational intensity by performing more arithmetic per byte fetched. Coalesced global reads and shared-memory reuse keep the kernel bandwidth-efficient.

hw3 (Blocked Floyd-Warshall)

The blocked Floyd-Warshall implementation uses cache blocking to improve locality and reduce cache misses. By operating on tiles, it keeps working sets in shared memory (GPU) or cache (CPU), cutting global memory access and improving arithmetic intensity across phases. The structure also enables parallel execution with predictable memory access patterns.

Build & Run

• Makefile-based builds are provided in hw1_odd-even-sort/, hw2_mandelbrot-set/, and hw3/ (use make -C <dir>).
• CUDA targets require nvcc and a compatible NVIDIA toolchain.
• MPI/OpenMP builds assume mpicc/mpicxx and OpenMP-enabled compilers in the environment.