Gracjan Adamus

Project Title: Exploration and Evaluation of non-LHC GPU Benchmarks for HEPScore4GPU.
Project Report: link
Summer Student Session Presentation: link

Selected from more than 14000 applicants, I completed a research project focused on GPU benchmarking within the IT Department, Computing for Experiments group, Worldwide LHC Computing Grid section (IT-CE-LCG).

My work involved containerizing HPC applications from various scientific domains (SPECFEM3D Globe, Quantum ESPRESSO, MILC QCD), collecting benchmark data with the HEP Benchmark Suite, and validating their reproducibility across multiple runs and on different machines (NVIDIA A100, V100, L4, T4). I also created and tested containers for SPECFEM3D Globe targeting AMD GPUs as well as CPU-only execution via MPI. I profiled the workloads with Nsight Systems and Nsight Compute. Additionally, I contributed some bug fixes and small features to HEP Benchmark Suite and HEPScore.

I presented my findings to fellow summer students and at group and section meetings. I also attended various CERN seminars and workshops, where I met notable figures such as Bjarne Stroustrup and Tim Mattson.

I worked on extending a Qt application for the client's needs, writing both QML widgets and C++ modules. I also took part in testing, code reviews, and writing technical documentation. I discovered and fixed numerous bugs and created a CI pipeline for building and publishing the artifacts. On my own initiative, I built a Python TUI for real-time mocking of the data received by the app, which resulted in much faster and easier development and QA reviews.

I was a part of an interdisciplinary team, spanning physicists and software developers. My main line of work was involved analysing space radiation data from monitors aboard ESA's satellites (IREM and RADEM). All my work was done in Python.

I created automation scripts for parsing 100s of CDF files into HDF5 format. I developed algorithms for automatic detection of space events like radiation belt passages, solar energetic particles and Forbush decreases with NumPy and pandas. I prepared a series of exemplary Jupyter Notebooks showcasing the algorithms, and providing analysis of said events. I also begun the process of porting the functions to be JIT-compilable with Numba.

nanobind_json serves as a seamless bridge for nlohmann::json objects with nanobind. A single header include is all you need to pass JSON data between Python and C++. The library is mentioned in the official nanobind documentation after a small email exchange with Wenzel Jakob.

I have taken over maintenance of the library, and fixed all issues to make it compatible with current nanobind versions. I also added tests and CI workflows. The library is built and tested on Linux, Windows, and macOS with their respective default compilers (gcc, MSVC, clang).

I participated in bi-weekly meetings of teams at University of Illinois Urbana-Champaign and AGH University of Kraków. I was responsible for porting our library's C++ to Python bindings from pybind11 to nanobind, and switching to cibuildwheel for our wheel building procedure on the Github Actions CI. I developed an input checker, which helps to guard the users from making simple mistakes like typos. I also fixed the pdoc documentation, with the most hacky piece of Python code I have ever written.