Comparative data analysis between experiment, HPC and neuromorphic systems
This project is a collaboration of scientists working in four HBP Subprojects:
- Neuroinformatics Platform (SP5)
- Brain Simulation Platform (SP6)
- HPAC Platform (SP7)
- Neuromorphic Computing Platform (SP9)
Sonja Grün (SP5/SP9) leads this project from a scientific point of view, Bernd Schuller (SP7) is responsible for the technical part.
A NEST simulation of a neural network model and the simulation of the same model on SpiNNaker (SP9) ideally lead to the same results. This use case aims at implementing a workflow that starts a simulation from the Collaboratory on both backends, i.e. a NEST simulation on an HPC system and another simulation of the same model on SpiNNaker, and that analyses and compares the results with the analysis tool Elephant (formerly SP5, now SP9). This way it can be evaluated for the first time if the NEST-HPC and the NM implementations lead to similar or the same results when simulating the same model of a neural network. A second step will be to also compare experimental data with the simulation results. The implementation of this use case will happen in two phases:
A NEST-Elephant demo organised by members from SP5 and SP6 at the HBP Summit 2015 successfully showed how a NEST simulation and a subsequent analysis of the results with Elephant could be started through the Collaboratory. However, simulation and analysis were not yet running on a system like HPC or neuromorphic (NM) hardware. Members from SP7 presented in another demo at the same Summit how a job on an HPC system could be started from the Collaboratory through the integrated UNICORE Portal. The goal of this project is to combine these two demos, i.e. to implement as a first use case a workflow that allows to start a NEST simulation through the Collaboratory on different backends (HPC or NM) and to analyse the results with Elephant.
The following demo was shown as part of HPAC presentation during the HBP Platform Release on 30th March 2016:
Problems to be solved
The following problems require special attention and need to be solved in co-design with the users for implementing the use case successfully.
- Provenance tracking of the entire, complex workflow is needed in order to enable reproducibility.
- The results of the neuromorphic simulation and the experimental datasets need to be transferred to the HPC storage and/or to the Collab storage for the analysis step with Elephant.
- The simulation and analysis results need to be visualised, which may require downloading the data from the HPC storage to the user’s computer.
- A seamless integration with the Collaboration is required to start the workflow from this portal, which in particular requires submitting and managing HPC jobs.
- The analysis tool Elephant is not parallelised yet. Therefore, using a parallel programming framework to enable (trivial) parallelisation of advanced analysis routines would allow the analysis of significantly larger datasets directly on a supercomputer.