Contributor: Isaac Morales Santana
Mentors: Vassil Vassilev and Lukas Breitweiser.
Organization: CERN-HSF
Project repo: https://github.com/BioDynaMo/biodynamo
This project took place during Google Summer of Code 2024, with the support from CERN and Compiler Research team.
BioDynaMo is an agent-based simulation platform designed to facilitate complex simulations, particularly in fields like cancer research, epidemiology, and social sciences. It leverages ROOT—a framework widely used in high-energy physics—for statistical analysis, random number generation, C++ Jupyter notebooks, and I/O operations. However, enhancing BioDynaMo’s performance remains a key challenge.
BioDynaMo’s reflection system, which utilizes Cling (an interactive C++ interpreter from ROOT), experiences significant runtime performance and memory usage issues. The repeated parsing of library descriptors by Cling introduces inefficiencies that slow down the startup phase and consume excessive memory. These bottlenecks are especially evident in simulations with a low number of time steps, as a substantial portion of the time is spent on parsing rather than on actual computations.
The primary goal of the GSoC project was to integrate ROOT’s C++ Modules into BioDynaMo to minimize these performance issues. C++ Modules offer an efficient on-disk representation of C++ code, reducing the need for repeated parsing of invariant code. By implementing these modules, the project aimed to optimize runtime memory usage and improve overall performance
- Reworking CMake Rules: The project incorporated ROOT and another packages efficiently using FetchContent, modifying CMake rules accordingly (e.g., PR #365 and #387, both merged)
- Replacing genreflex with rootcling: This switch was crucial to enable C++ Modules and streamline the generation of reflection information (PR #379)
- C++ Modules changes Among other things, I used automatic generation for the module map with relative paths,
modified the
selection.xmlfile to support the new dictionaries and fixed headers with missing includes (PR #385. - Updated some CI workflows I fixed some failing workflows in PR #377. Also, I did some minor changes in some flags in PR's #378 and #367
The results have been promising, showcasing significant performance gains. Benchmarking revealed
improvements ranging from 18% reduction in peak memory usage with the default modules.idx to 25% with the
updated one.
Moreover, the startup phase saw an impressive 80% reduction in time, thanks to the optimized
handling of header parsing. That highlights the efficiency of C++ Modules in minimizing
Cling’s parsing overhead
As expected, the simulation time did not show an appreciable improvement. However, in the unit tests, the time was 33% lower. I believe this is because unit tests involve a lot of parsing and Cling calls.
Despite these advances, several challenges remain. PR's #365 is ready to merge and #385 needs some changes. For instance, memory leaks have been observed when using the new
ROOT_GENERATE_DICTIONARY, even with C++ Modules disabled. Additionally, the build system for individual demos has
caused compatibility issues with the main build system. Also, there is a problem with the Jupyter notebooks.
Resolving these issues and finalizing the integration of C++ Modules will be essential for ensuring long-term stability and reliability.
Looking ahead, further optimizations are planned, including potential module-based optimizations for BioDynaMo’s core components. Collaboration with the BioDynaMo team continues, with upcoming meetings scheduled to align efforts and resolve outstanding issues.
The integration of C++ Modules has proven effective in reducing memory usage and startup time, although some hurdles remain. Continued collaboration and testing will be crucial to fully realize the performance potential of BioDynaMo, enabling more efficient simulations for researchers in computational biology.
First of all, I would like to express my sincere gratitude to Google Summer of Code for giving me the opportunity to work on this project. I am deeply grateful to my mentors, Vassil Vassilev and Lukas Breitweiser, for their invaluable guidance, patience, and support throughout the project. Their expertise and mentorship were instrumental in overcoming challenges and pushing the project forward.