XPU-Point

A toolkit for profiling and performance evaluation of heterogeneous CPU-GPU applications across NVIDIA and Intel GPU systems.

Overview

XPU-Point provides the necessary tools and scripts to analyze heterogeneous applications through three main components:

1. XPU-Profiler - Profiles heterogeneous applications and collects basic block vectors for similarity analysis
2. XPU-Timer - Evaluates performance of selected regions
3. Performance Extrapolation - Extrapolates performance results and generates visualization plots

This repository includes an example benchmark (GROMACS) to demonstrate the end-to-end methodology.

Quick Start

Prerequisites

Hardware Requirements:

• Linux x86 system with NVIDIA GPU
  • CUDA version: >= 8.0 && <= 11.x
  • CUDA driver version: <= 495.xx
• Linux x86 system with Intel GPU
• At least 500 GB of free disk space

Software Dependencies:

• Docker
• NVIDIA drivers
• Intel drivers
• CUDA
• oneAPI

Installation

1. Clone the repository:

git clone https://github.com/nus-comparch/xpupoint && cd xpupoint

2. Set up the environment using Docker:

# Build the docker image
make docker.build

# Run the docker image  
make docker.run

# Compile XPU-Point tools and benchmarks
make

Running Experiments

1. Navigate to the one of the benchmark directories (for example, GROMACS):

cd benchmarks/gromacs

2. Run XPU-Point analysis:

# Run all test cases (NOTE: The full test suite can take a long time to complete)
./run-xpupoint all

# Or run individual tests (--help shows the available tests)
./run-xpupoint <test_directory_name>

3. Generate visualization:

./make-graphs

How It Works

The run-xpupoint script automates the complete analysis workflow:

1. XPU-Profiler identifies representative regions in the application
2. XPU-Timer measures performance of both the full application and individual representative regions
3. Results are processed and can be visualized as tables or graphs

System Requirements

Component	Requirement
Programs	C++ programs, Python/Shell scripts
Compilation	CUDA, oneAPI, Make, GCC
Binaries	Pin 3.30, GTPin 4.5.0, NVBit 1.5.5
Runtime	NVIDIA and Intel GPU Drivers
Hardware	NVIDIA GPU systems and Intel GPU systems
Metrics	Cycles, RDTSC, Runtime
Disk Space	~500 GB
Setup Time	~1 day
Experiment Time	~1 week

Verification

Before running experiments, verify your GPU platforms are functioning:

# For NVIDIA GPUs
nvidia-smi

# For Intel GPUs  
sycl-ls

Important Notes

Performance Considerations:

• Results are closely tied to the specific execution environment
• XPU-Timer results may be affected by other jobs running on the same machine
• Profiling runs for larger applications can take several hours

Availability

• GitHub Repository: nus-comparch/xpupoint
• Archived Version: Zenodo DOI: 10.5281/zenodo.16801115
• License: Open source (see repository for specific license details)

Contributing

This project is publicly available and contributions are welcome. Please check the repository for contribution guidelines.

Support

For issues and questions, please use the GitHub Issues section of this repository.