CERN SFT Gsoc 2016 ideas page

Google Summer Of Code 2016

We participated in the Google Summer of Code with success each year since 2011, so we are applying again in 2016!
We intend to offer three options, as soon as the project ideas will be ready: pick one of the Project ideas which are grouped into categories, have a look to the section dedicated to 'Blue sky' or propose your own great idea for this summer: we are looking forward to hearing about your perspectives. A list of our mentors (and their areas of expertise) can be found here.

We encourage students who plan to apply to contact us  about their interest and explain their project ideas as early as possible. Our experience from our previous GSoC participation was that frequently an initial student application either needs to be reworked in close collaboration with a future mentor, or at least can benefit from feedback and discussion. Please do not forget to provide us with some relevant information about yourself (for example CV, past projects, personal page or blog, linkedin profile, github account, etc.).

Before submitting an application please consult the official GSoC FAQ where you can find some good advice on writing a successful application. The application should be submitted through the GSoC Web page before the 25th of March.

Project ideas

Geant Vector Prototype and Geant4 Simulation Toolkit

Every LHC experiments is a large scale user of simulation, involving the elementary particles created by the LHC beam collisions and their path through its detectors. Almost half the CPU resources of each experiment (around 100,000 cores) are used constantly to produce simulated events. This is now done using the Geant4 toolkit, the current production detector simulation toolkit for High Energy Physics (HEP) experiments.  R&D is being undertaken in the GeantV project into a new generation of detector simulation, seeking to use existing 'large core' hardware more efficiently, in order to meet the rapidly rising experiment simulation needs, and to be well adapted to other existing and planned architectures, including CPUs with larger vector registers and accelerators ( such as the Intel Xeon Phi & GPUs). The code required to model diverse types of particles and interactions, and to model the complex geometries of detectors is large. Due to this it overwhelms the caches of current CPUs, significantly reducing the efficiency of utilisation on today’s hardware. This effort is focused on identifying ways to reorganise the work so that more data (e.g. multiple particles or rays) is processed by each function. By reusing the code on 'nearby' data we aim to use the memory architectures of today’s hardware better. At the same time we prepare the way to obtain good performance on tomorrow’s hardware.

Geant4 and GeantV are open source, developed by physicists and engineers from laboratories and universities around the world.  Developments are ongoing to improve its computing and physics precision, its scope of application, and to better utilize current and emerging computer architectures. The simulation programs of the LHC experiments are in constant large scale use, and the total number of simulated events produced is becoming a growing limitation in the analysis potential for some interesting types of new physics.

As a result the goal of the project is to explore different ways to reduce the execution time on today’s complex commodity CPUs, and to prototype how to use it efficiently on the many-core hardware of the future (tens, hundreds of cores, threads or ‘warps’). The code required in Geant4 to model diverse types of particles and interactions, and to model the complex
geometries of detectors spans hundreds of classes and tens to hundreds of thousands of lines of code. Due to this it overwhelms the caches of current CPUs, significantly reducing the efficiency of utilisation on today’s hardware. GeantV addresses these challenges by transporting together bunches of particles, in order to profit from the structure and parallelism in modern hardware, exploiting cacahce, vector instructions and multiple cores using multi-threading (MT).

Implementation of task-based transport for GeantV

Description: The current parallelism model of GeantV is data-oriented: a static set of threads handles its work, fetching one "basket" of tracks and transporting each track for a step. Threads are fed with work from a common input concurrent queue. After an injection of a set of initial tracks, further work is generated by the transport threads, which create new baskets at the end of each step. Each step can also generate a trace (or detector 'hit'), which is proto-data for output. Other work will also go on, summarising the detector hits into detector summaries ('digits'), and performing its output.  We are seeking to adapt the steering of this work to a task-based approach, preferably using Thread Building Blocks (TBB), to profit the flexibility of this approach, and from the facilities provided by its powerful library.  A first TBB implementation exists based on an initial version of the GeantV scheduler, which serve as a starting point and potential inspiration.

Task ideas and expected results:  

  • Evaluate the existing TBB-based implementation by identifying and understanding its current bottlenecks, and comparing it to the 'static' thread approach. Expected outcome is a refined task-based version, with improved performance from addressing the most important bottlenecks.
  • Evaluate different approaches for structuring the tasks. Innovative ideas will be welcome. Expected outcome is a design and partial implementation of a second task-based implementation targeting improved performance.

Requirements: Strong C++ skills, experience with parallel programming required. Experience in using TBB or other task-based threading libraries will be considered a strong advantage. Knowledge in the field of physics is a plus, but not a requirement.

Mentor: Andrei Gheata

Web page:

Source code:

New methods for integrating trajectory in field

Geant4 and GeantV use Runge-Kutta methods  to integrate the motion of charged particles in a non-uniform electromagnetic field.  Methods must provide good integration accuracy for the integration and to cost a minimum of computation time.  Integration is used to identify the intersection point between the curved track and the volume boundaries.  Due to the large number of steps and the cost of the evaluations of the field, the integration and intersection are a performance critical part of detector simulation. Recent work has introduced new RK methods which reduce the number of field evaluations required, and has the potential to decreased the computation time.

Task ideas:

  • Introduce multi-step integration methods and compare their performance with the available RK methods, or
  • Introduce heuristics for choosing an appropriate RK method, using knowledge of the accuracy requirements and the length of the current step, and the characteristics of the magnetic field.

Expected results: Working implementation of a) one or more multi-step integration methods inside Geant4 and/or Geant4 module for tracking in field, or b) integration methods which combine RK methods of different order for improved performance.

Requirements: This project requires prior exposure to Numerical Analysis and familiarity with either C++, C or Java programming.  Exposure to either numerical methods for solving Ordinary Differential equations (ODEs) or tools for analysing data such as ROOT or R will be valuable. Both programming skill and knowledge of numerical methods for ODEs will be improved by undertaking this project.

Mentor: John Apostolakis​
Web site:
Source code




The ROOT system ( provides a set of OO frameworks with all the functionality needed to handle and analyze large amounts of data in a very efficient way. Having the data defined as a set of objects, specialized storage methods are used to get direct access to the separate attributes of the selected objects, without having to touch the bulk of the data. Included are histogramming methods in an arbitrary number of dimensions, curve fitting, function evaluation, minimization, graphics and visualization classes to allow the easy setup of an analysis system that can query and process the data interactively or in batch mode, as well as a general parallel processing framework, PROOF, that can considerably speed up an analysis. Thanks to the built-in C++ interpreter the command language, the scripting, or macro, language and the programming language are all C++. The interpreter allows for fast prototyping of the macros since it removes the, time consuming, compile/link cycle. It also provides a good environment to learn C++. If more performance is needed the interactively developed macros can be compiled. ROOT's new C++11 standard-compliant interpreter is Cling, an interpreter built on top of Clang ( (link is external)) and LLVM ( (link is external)) compiler infrastructure. Cling is being developed at CERN as a standalone project. It is being integrated into the ROOT data analysis ( framework, giving ROOT access to an C++11 standards compliant interpreter. ROOT is an open system that can be dynamically extended by linking external libraries. This makes ROOT a premier platform on which to build data acquisition, simulation and data analysis systems. ROOT is the de-facto standard data storage and processing system for all High Energy Phyiscs labs and experiments world wide. It is also being used in other fields of science and beyond (e.g. finance, insurance, etc).

Enhance C-Reduce to work with ROOT



C-Reduce (, is a tool which aims to reduce bug reports. It transforms user's source files to make them as minimal as possible. Minimalistic bug reproducers are easy to debug and convert into regression tests. C-Reduce is fairly mature and well adapted to minimize crashes in compilers. The project will be mainly focused towards making C-Reduce easier to use with ROOT and it's interactive C++ interpreter cling.


Expected results: Extend C-Reduce to be able to reduce easily ROOT bug reports. Optionally extend C-Reduce to minimize ROOT's data files. Implement tests for all the realized functionality. Prepare a final poster of the work and be ready to present it.

Required knowledge: Intermediate level of C++, some experience with Clang

Mentor:Vassil Vassilev

Extend clad - The Automatic Differentiation

Description: In mathematics and computer algebra, automatic differentiation (AD) is a set of techniques to numerically evaluate the derivative of a function specified by a computer program. Automatic differentiation is an alternative technique to Symbolic differentiation and Numerical differentiation (the method of finite differences). Clad ( is based on Clang which will provides the necessary facilities for code transformation. The AD library is able to differentiate non trivial functions, to find a partial derivative for trivial cases and has good unit test coverage. There was a proof-of-concept implementation for computation offload using OpenCL.

Expected results:The student should teach AD how to generate OpenCL/CUDA code automatically for a given derivative. The implementation should be very well tested and documented. Prepare a final poster of the work and be ready to present it.

Required knowledge: Advanced C++, Clang abstract syntax tree (AST), CUDA/OpenCL basic math.

Mentor:Vassil Vassilev

Interactive features for the JSROOT web geometry viewer

Description: The ROOT project is developing a JavaScript library reading and rendering ROOT objects (1D, 2D, 3D histograms, 1D & 2D graphs, multi-graphs, 3D geometries) in any modern web browser. One part of this task is to render 3D objects (like detector geometries) using the THREE.js library. This is advancing rapidly and most of the shapes have been implemented. Nevertheless, many interactive features are missing, such as navigation/selection of volumes (picking) and introducing/moving cutting plane. This has to be implemented. And the current code can handle and display several hundred of volumes, but typical detector geometries can be composed of several millions of volumes. The code should be able to deal with such big geometry (memory and performance wise). See the JSROOT page and actual status here.

Expected results:

  • Provide more interactive features to web geometry viewer
  • Performance optimization for large geometries


Required knowledge:Good knowledge of JavaScript. Experience with 3D computer graphics. Knowledge of C++ would be an asset

Mentor:Bertrand Bellenot and Sergei Linev


TMVA Project in Machine Learning

Description: Toolkit for Multivariate Analyses (TMVA) is a machine-learning framework integrated into the ROOT software framework, containing ML packages for classification and regression frequently used by high-energy physicists in searches for new particles, used for example in the discovery of the Higgs Boson. Recently TMVA has been undergoing a significant makeover both in performance, features and functionality.

There are a number of possible areas of contribution, for example:

Task ideas

  • Improvement of memory management and data-handling for parallel running

  • GPU support for intensive deep learning training applications

  • Interfaces to other machine-learning tools

  • Support for multi-objective regression

  • Support for feature engineering

Expected Results: working implementation of these features in TMVA leading to improved software performance

Requirements: Strong C++ background is desired, strong machine learning knowledge is a plus. 

GitHub repositoryTMVA

Mentors: Sergei Gleyzer​, Lorenzo Moneta

Integrating Machine Learning in Jupyter Notebooks

Description: Improving user experience with ROOTbooks and TMVA. A ROOTbook is a ROOT-integrated Jupyter notebook. ROOT is a software framework for data processing, analysis, storage and visualization. Toolkit for Multivariate Analysis (TMVA) is a ROOT-integrated package of Machine Learning tools. Jupyter notebook is a web-based interactive computing platform that combines code, equations, text and visualizations.

Task ideas:

  • Integrate a list of features, currently available in the TMVA Graphical User Interface, into the ROOTbook environment. This includes Receiver Operating Characteristic (ROC) curves, feature correlations, overtraining checks and classifier visualizations.
  • Extend the ROOT-Python binding (or PyROOT) for the use of TMVA in ROOTbooks. This includes simplifying parameter specification for booking and training classifiers, improving output readability and code clarity.
  • Implement interactive training mode in the ROOTbook environment.
  • Interactive feature engineering with JavaScript visualization.
  • Interactive deep learning optimization.

Expected results: working implementation of the TMVA-ROOTbooks integration layer.

Requirements: Python, C++, Javascript, machine learning, familiarity with notebook technology is a plus

Links: TMVA, ROOTbooks

Mentors: Sergei Gleyzer​ and  Enric Tejedor


Reflection-based Python-C++ language bindings: cppyy

cppyy is a fully automated, run-time, language bridge between C++ and Python. It forms the underpinnings for PyROOT, the Python bindings to ROOT, the main persistency and analysis framework in High Energy Physics (HEP), is used to drive the frameworks of several HEP experiments, and is the environment of choice for analysis for many HEP physicists. cppyy is the only Python-C++ bindings technology that can handle the scale, complexity, and heterogeneity of HEP codes. There are two implementations, one for CPython, and one for PyPy.

Source codes, documentation, and downloads:  and

Both the CPython and PyPy implementations support the CINT and Reflex reflection systems, the CPython version also supports Cling, which is based on Clang/LLVM. The goal is to move both to Cling on a code base that is as much shared as possible.


Integrate the Cling backend into PyPy/cppyy

Description: Cling, being based on Clang/LLVM can parse the latest C++ standard (C++11/C++14). A Cling backend exists for CPython/cppyy, but not yet for PyPy/cppyy. A common backend could serve both projects, and would reduce the cost of new features, making them much quicker available.

Expected results: Implement a Cling backend on libCling directly, using the CPython implementation as a starting point, for use by both CPython and PyPy. Package this backend for distribution. Design and implement a method for distribution of Clang modules with the standard Python distribution tools.

Requirements: Working knowledge of C++, good knowledge of Python

Mentor: Wim Lavrijsen

Sixtrack numerical accelerator simulation​

SixTrack is a software for simulating and analysing the trajectory of high energy particles in accelerators. It has been used in the design and optimization of the LHC and is now being used to design the High-Luminosity LHC (HL-LHC) upgrade that will be installed in the next decade. Sixtrack has been adapted to take advantage of large scale volunteer computing resources provided by the LHC@Home project. It has been engineered to give the exact same results after millions of operations on several, very different computer platforms. The source code is written in Fortran, and is pre-processed by two programs that assemble the code blocks and provide automatic differentiation of the equation of motions. The code relies on the crlibm (link is external) library, careful arrangement of parenthesis, dedicated input/output and selected compilation flags for the most common compilers to provide identical results on different platforms and operating systems. An option enables the use of the Boinc (link is external) library for volunteer computing. A running environment SixDesk is used to generate input files, split simulations for LHC@Home (link sends e-mail) or CERN cluster and collect the results for the user. SixTrack is licensed under LGPLv2.1.

A strong background in computer science and programming languages as well the interest to understand computational physics methods implemented in the code are sought. The unique challenge will be offered to work with a high-performance production code that is used for the highest energy accelerator in the world - and thus the code's reliability and backward compatibility cannot be compromised. There will be the opportunity to learn about methods used in simulating the motion of particles in accelerators.


Optimize and Integrate Standalone Tracking Library (SixTrackLib)

Description: Benchmark a standalone tracking library in C targeting both CPU and GPU and integrate in SixTrack. The inner loop uses a simple data structure based on contiguous arrays that can be generated by SixTrack or external programs and can be hosted in the CPU or GPU main memory. In case of GPU, the ideal number of particle per core (even one such that coordinates do not leave internal registers) and kernel size should be evaluated for speed.

Expected results: Running code which rely only on the newly rewritten library to perform tracking simulations and test suite that proves that old and new implementation produce identical results.

Mentors: Ricardo De Maria (link sends e-mail)

Requirements: Experience with Fortran, C, OpenCL, calculus and a background in physics. 

New physics models

Description: Implement, test and put in production a new solvers for exact bending dipoles, combined function magnets and radiation effects.

Expected results: The user can build accelerator simulations with more accurate models for low energy machines and/or machines with radiation effects.

Mentors: Ricardo De Maria (link sends e-mail)

Requirements: Fortran, calculus, accelerator physics.


The CERN Beam Longitudinal Dynamics code, BLonD, is used to simulate the dynamics of particle bunches in synchrotrons. It contains a vast range of different physics features to model multiple-harmonic RF systems, feedback loops, and collective effects, and has been applied for many studies and several machines in- and outside of CERN. Whether it is to understand previously unexplained observations, or whether it is to predict and optimize parameters for future, simulations often require multi-bunch modelling with millions of particles and calculations of collective effects (in frequency or time domain), sometimes over millions of iterations, which can make the simulations computationally very expensive.

BLonD Code Optimisation

Description: The code was originally written in python. In order to significantly reduce the runtime, it will be translated to C++ and algorithms are going to optimized by a BLonD developer during the coming year. This will not only require a complete rewriting, but also a major restructuring of the code, where creativity, initiative, and latest technologies will be needed. As a Gsoc student, you could contribute to explore different parallelization techniques on CPUs and GPUs for different parts of the code, as well as different data structures and overall software architecture options to increase computational efficiency.

Expected results: Determine the best architecture and parallelization option(s) for the BLonD code.

Requirements: Strong skills in C++ and parallelization techniques. Some experience with python would be of advantage. Furthermore, a minimal physics background that allows for understanding the underlying equations.

Mentors: Helga Timko and Danilo Piparo

Website (with links to source code & documentation):


'Blue sky' ideas

Improvement of the VDT Mathematical Library

The VDT (link is external) mathematical library is a collection of optimised, inline and vectorisable mathematical functions. Its adoption allowed to reduce remarkably the runtime of the data processing workflows of CMS experiment at the LHC.

This project aims to further expand the functionality of the VDT (link is external) mathematical library. Two main areas can be explored, namely:

1. Integration with OMP4 and support for simd vectors

The VDT functions can be enhanced in order to support the OpenMP4 programming interface relative to vectorisation. In additon, by templating the VDT (link is external) functions, the explicit vectorisation through the usage of array types such as the gcc and clang built-in types or the Vc array types.

2. Integration of existing limited precision/domain function implementations

Often the usage of a certain mathematical function requires the support of a limited domain or a limited precision. This activity aims to complement the existing VDT (link is external) functions implementations with others characterized by a reduced precision or input range. An appropriate formulation of the interfaces of these functions has to be adopted, for example adopting generic programming principles through the usage of templates.

Mentors: Vincenzo Innocente, Danilo Piparo


Here is the list of our mentors and their areas of expertise:

Contact information

Please do not hesitate to contact us if you are planning to apply for any of the above projects: