NEGATIVA_ML

---

This repository provides the open-source implementation of our paper “The Hidden Bloat in Machine Learning Systems.”

The project includes the components described in the paper, the kernel detector and kernel locator. The compaction component is not included in this release and will be made public upon acceptance of our companion paper.

Overview

NEGATIVA_ML analyzes machine learning (ML) workloads to identify unused GPU code segments in shared libraries. Given an ML workload, negativa_ml outputs:

• A list of shared libraries loaded during execution.
• The unused GPU code segments within those libraries.

See the Quick Start section below for usage instructions.

---

Installation

Tested environment: Ubuntu 20.04

Requirements: CUDA must be properly installed and configured. Refer to the official CUDA installation guide: CUDA Installation Guide for Linux.

Step-by-step setup

1. Install Rust: https://rust-lang.org/tools/install/

2. Install dependencies:

   sudo apt update
   sudo apt install libspdlog-dev cmake

3. Set LD_LIBRARY_PATH:

   export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/extras/CUPTI/lib64

4. Run tests:

   make test

   All tests should pass successfully.

5. Install and verify:

   make install
   negativa_ml --help

---

Quick Start

Example ML workloads are available in the examples directory. Here’s how to analyze the demo example:

1. Build the demo:

   cd examples/demo && ./build.sh

   The demo example compiles a shared library libdemo.so and a main executable main that uses it. The shared library contains two GPU kernels: matrixMulGPU and setScalarItems.

2. Verify the demo:

   cd build && ./main matmul

   This should output SUCCESS.

3. Run NEGATIVA_ML analysis:

   negativa_ml debloat -- $PWD/main

   Note: The executable path must be absolute.

4. View the results:

   The analysis results are stored in the nml_workspace directory:

   nml_workspace
   ├── trace.json        # Traced kernels and loaded shared libraries
   ├── spans/            # Located unused GPU code segments
   │   └── libdemo.so.json  # Lists unused code segments in `libdemo.so`
   

   The spans directory contains the unused GPU code segments for each shared library. These can be used as input for the compaction component (not yet released).

---

Reconstructing Shared Libraries

You can verify the detected code segments are unused by modifying them and rerunning the workload. The following command reconstructs a shared library by overwriting unused code regions with 0x1:

negativa_ml reconstruct \
  --span-path nml_workspace/spans/libdemo.so.json \
  --output-dir ./reconstructed

This will produce a reconstructed version of the shared library in ./reconstructed/. You may replace the original shared library with this version to verify correctness. Remember to back up the original file first.

For convenience, a helper script debloat.sh is provided under the demo example to automate this process.

---

Command Line Usage

The main executable is negativa_ml, which supports four subcommands:

Command	Description
trace	Traces kernel launches and loaded shared libraries during the workload execution.
locate	Identifies unused GPU code segments in shared libraries based on trace results.
debloat	Runs trace and locate sequentially, producing final analysis results.
reconstruct	Rebuilds shared libraries with unused code segments set to 0x1.

---

Citation

If you use this code in your research, please cite our paper:

@inproceedings{zhanghidden,
  title={The Hidden Bloat in Machine Learning Systems},
  author={Zhang, Huaifeng and Ali-Eldin, Ahmed},
  booktitle={Eighth Conference on Machine Learning and Systems (MLSys)},
  year={2025}
}