Bachelor's Thesis: "Exploring Fuzzy Tuning Technique for Molecular Dynamics Simulations in AutoPas"

This repository contains all the collected data, scripts, and the final thesis document for my Bachelor Thesis at the Technical University of Munich (TUM). The thesis was written at the Chair of Scientific Computing under the supervision of Manish Kumar Mishra and Samuel Newcome.

The thesis explores a novel tuning technique for molecular dynamics simulations in the AutoPas framework. The technique is based on fuzzy logic and aims to predict close-to-optimal simulation configurations from a set of live data collected during the simulation. The goal of this thesis is to evaluate the feasibility of the fuzzy tuning technique and to compare it with existing tuning techniques.

Abstract

AutoPas is a high-performance, auto-tuned particle simulation library for many-body systems, capable of dynamically switching between algorithms and data structures to guarantee optimal performance throughout the simulation. This thesis introduces a novel fuzzy logic-based tuning strategy for AutoPas, allowing users to guide the tuning process by specifying custom Fuzzy Systems, which can be used to efficiently prune the search space of possible parameter configurations. Efficient tuning strategies are crucial, as they allow for discarding poor parameter configurations without evaluating them, thus reducing tuning time and improving overall library performance.

We demonstrate that a data-driven approach can automatically generate Fuzzy Systems that significantly outperform existing tuning strategies on specific benchmarks, resulting in speedups of up to 1.96x compared to the FullSearch Strategy on scenarios included in the training data and up to 1.35x on scenarios not directly included.

The Fuzzy Tuning Strategy can drastically reduce the number of evaluated configurations during tuning phases while achieving comparable tuning results, making it a promising alternative to the existing tuning strategies.

Table of Contents

1. Introduction
2. Theoretical Background
   • Molecular Dynamics
     • Quantum Mechanical Background
     • Classical Molecular Dynamics
     • Potential Energy Function
     • Numerical Integration
     • Simulation Loop
   • AutoPas
     • Autotuning in AutoPas
     • Tunable Parameters
     • Tuning Strategies
   • Fuzzy Logic
     • Fuzzy Sets
     • Fuzzy Logic Operations
     • Linguistic Variables
     • Fuzzy Logic Rules
     • Fuzzy Inference
3. Implementation
   • Fuzzy Logic Framework
   • Rule Parser
   • Fuzzy Tuning Strategy
     • Component Tuning Approach
     • Suitability Tuning Approach
4. Proof of Concept
   • Data-Driven Rule Extraction
     • Decision Trees
     • Conversion of Decision Trees to Fuzzy Systems
   • Fuzzy Systems for md flexible
     • Data Collection
     • Data Preprocessing
     • Component Tuning Approach
     • Suitability Tuning Approach
5. Comparison and Evaluation
   • Exploding Liquid Benchmark (Included in Training Data)
   • Spinodal Decomposition MPI (Related to Training Data)
   • Further Analysis
     • Quality of Predictions During Tuning Phases
     • Optimal Suitability Threshold
     • Generalization of Rule Extraction Process
6. Future Work
   • Dynamic Rule Generation
   • Improving Tuning Strategies
   • Simplification of the Fuzzy System to Decision Trees
7. Conclusion

Thesis

The thesis is available in LaTeX format in this repository. You can access the rendered version in PDF format by clicking the following link:

Read the Thesis (PDF)

Slides

The presentation slides are available in PDF format and can be accessed via the following link:

View the Slides (PDF)

Quick Guide for Creating New Rule Sets

This section provides a quick guide on how to create new rule sets similar to the ones developed in this thesis. ( Suitability.frule and Components.frule)

Tip

1. Collect Data (Logfiles of AutoPas Simulations):

   • To enable logging make sure to enable: -DAUTOPAS_LOG_TUNINGDATA=ON -DAUTOPAS_LOG_LIVEINFO=ON -DAUTOPAS_LOG_TUNINGRESULTS=ON or similar in the CMake configuration.

   • -DMD_FLEXIBLE_PAUSE_SIMULATION_DURING_TUNING=ON is also recommended to guarantee fair comparisons between different configurations.

   • The script data/CoolMuc2_launch_DataCollection.py provides a template for launching such simulations on the CoolMuc2 cluster.

   • You should end up with a bunch of log files for each evaluated scenario consisting of: AutoPas_tuningData_*.csv, AutoPas_tuningResults_*.csv, AutoPas_iterationPerformance_*.csv and Scenario_*.out files. Copy them to your local machine and place them in the data/new folder.

   • If you want to cleanup the *.out files, you can use the script data/remove_empty_lines.sh script and run it in the data/new directory.

2. Sort Data into Folders:

   • Currently all logfiles are placed in the data/new directory, with only the timestamp as a hint for the actual scenario. To create meaningful datasets, we sort all logfiles of the same scenario into a separate folder. You can use the script data/extract_data.py to automatically attempt to sort the logfiles into folders based on the timestamps present in the *.out and *.csv files.
   • Check all the created folders manually and verify that all logfiles are correctly placed in the corresponding scenario folder.

3. Create Permanent Data Folders:

   • Copy the created folders in data/new to a permanent location in the data directory.

4. Create Rule Sets:

   • Go to the data-analysis directory and adapt one of the templates rule-extraction-components-template.ipynb or rule-extraction-suitability-template.ipynb to your needs.

     • You will need to adapt the PATH to the new data folders in the first cell of the notebook.

   • After running the entire notebook, you will find the extracted rule sets in the suitability-rules or components-rules directory.

   • Unfortunately, you need to manually create the output mappings in the *.frule files. You can look at Suitability.frule or Components.frule for reference. Tools such as Github-Copilot can help with this repetitive task.

   • The resulting *.frule files can now be used in AutoPas with the Fuzzy Tuning Strategy by adding --fuzzy-rule-filename {RULES}.frule to the command line arguments.

   • If you want you can format the rules to your liking and make them more readable.