This title appears in the Scientific Report :
2017
Please use the identifier:
http://hdl.handle.net/2128/15114 in citations.
Automatically generating HPC-optimized code for simulations using neural mass models
Automatically generating HPC-optimized code for simulations using neural mass models
High performance computing is becoming every day a more accessible and desirable concept for researchers in neuroscience. Simulations of brain networks and analysis of medical data can now be performed on larger scales and with higher resolution. However many software tools which are currently avail...
Saved in:
Personal Name(s): | Woodman, Marmaduke |
---|---|
Diaz, Sandra (Corresponding author) / Peyser, Alexander | |
Contributing Institute: |
Jülich Supercomputing Center; JSC |
Imprint: |
2017
|
Conference: | 26th Computational Neuroscience Meeting, Antwerp (Belgium), 2017-07-25 - 2017-07-25 |
Document Type: |
Poster |
Research Program: |
Human Brain Project Specific Grant Agreement 1 Deutschland - USA Zusammenarbeit in Computational Science: Mechanistische Zusammenhänge zwischen Struktur und funktioneller Dynamik im menschlichen Gehirn Supercomputing and Modelling for the Human Brain Theory, modelling and simulation Computational Science and Mathematical Methods SimLab Neuroscience |
Link: |
OpenAccess |
Publikationsportal JuSER |
High performance computing is becoming every day a more accessible and desirable concept for researchers in neuroscience. Simulations of brain networks and analysis of medical data can now be performed on larger scales and with higher resolution. However many software tools which are currently available to neuroscientists are not yet capable of utilizing the full power of supercomputers, GPGPUs and other computational accelerators. The Virtual Brain (TVB)[1] software is a validated and popular choice for the simulation of whole brain activity. With TVB the user can create simulations using neural mass models which can have outputs on different experimental modalities (EEG, MEG, fMRI, etc). TVB allows the scientists to explore and analyze simulated and experimental signals and contains tools to evaluate relevant scientific parameters over both types of data[2]. Internally, the TVB simulator contains several models for the generation of neural activity at the region scale. Most of these neural mass models can be efficiently described with groups of coupled differential equations which are numerically solved for large spans of simulation time. Currently, the models simulated in TVB are written in Python and have not been optimized for parallel execution or deployment on High Performance Computing architectures. Moreover, several elements of these models can be abstracted, generalized and re-utilized, but the design for the right abstract description for the models is still missing. In this work we want to present the first results of porting several workflows from The Virtual Brain into High Performance Computing accelerators. In order to reduce the effort required by neuroscientist to utilize different HPC platforms, we have developed an automatic code generation tool which can be used to define abstract models at all stages of a simulation. These models are then translated into hardware specific code. Our simulation workflows involve different neural mass models (Kuramoto [3], Reduced Wong Wang [4], etc ), pre-processing and post-processing kernels (ballon model [5], correlation metrics, etc). We discuss the strategies used to keep the code portable through several architectures but optimized to each platform. We also point out the benefits and limitations of this approach. Finally we show initial performance comparisons and give the user an idea of what can be achieved with the new code in terms of scalability and simulation times. AcknowledgementsWe would like to thank our collaborators Lia Domide, Mihai Andrei, Vlad Prunar for their work on the integration of the new software with the already existing TVB platform as well as Petra Ritter and Michael Schirner for providing an initial use case for our tests. The authors would also like to acknowledge the support by the Excellence Initiative of the German federal and state governments, the Jülich Aachen Research Alliance CRCNS grant and the Helmholtz Association through the portfolio theme SMHB and the Initiative and Networking Fund. In addition, this project has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No 720270 (HBP SGA1). |