hetsort_cuda_shellsort

Overview

Implementation of HETsort that leverages a parallel shellsort kernel optimized for NVIDIA GPUs, focusing on sorting high volumes of data in heterogeneous computing environments and optimizing device memory usage.

Prerequisites

• NVIDIA GPU
• CUDA Toolkit 10.0+
• GCC: Version 8.5.0+
• GNU Make: 4.2.1+
• Python 3.6+

Structure

• src/: Source files for the main application.
• profiles/: Directory for profile outputs.

Compilation

make build        # Compiles the source code into an executable named 'main'.
make run          # Builds (if necessary) and runs the application.
make clean        # Removes generated profile outputs.
make clean-sqlite # Removes SQLite files from profiles directory.

Usage

./main <method> <distribution> <arraySize> <deviceMemory> <iterations> <warmup> <checkSorted> <gpuCount> <seed>

Parameters:

• method: Sorting method to use.
• distribution: Type of data distribution.
• arraySize: Number of elements to sort.
• deviceMemory: GPU memory to utilize.
• iterations: Number of times to run the benchmark.
• warmup: Number of warmup runs before timing.
• checkSorted: Whether to check if the array is sorted post-operation.
• gpuCount: Number of GPUs to use.
• seed: Random seed for data generation.

Benchmarking

./benchmark.sh

This script automates the running of different configurations to facilitate a comprehensive performance analysis.

./generate_all.sh

This script utilizes the Python scripts generate_csv.py, generate_merge.py, and generate_nsys.py to process the results into csv format.

Academic Context

This project is conducted as part of a Master's thesis in Computer Science at University of Oslo (UiO).