In the context of the VINEYARD project, Neurasmus and ICCS have developed the BrainFrame framework in order to take advantage of heterogeneous ensembles inside datacentres. The framework’s first goal is to accelerate the simulation of large, complex neuronal networks, relevant to current neuroscientific research at unprecedentedly short simulation times through transparently using different accelerators. The second, but equally important, goal is to provide a familiar and easy frontend for the neuroscientists that hides the complexity of heterogeneous acceleration, without requiring the constant mediation of a hardware/software engineer for achieving acceleration.
The user (neuroscientist) initially sets up the characteristics of the neuronal network and decides on the simulation parameters through an intuitive Web GUI. Also, the user has the option to a priori determine the platform that the simulation will be executed on. The webserver then passes the user-provided info onto the PyNN framework, which makes the final choice of accelerator type based on a number of parameters (to be analysed later). Finally, when the simulation has ended, the results are sent back to the Web GUI.
The image is a screenshot of the web GUI, with the most important elements highlighted. The user / neuroscientist will:
- Select the platform that the simulation will run on: Xeon Phi, Maxeler DFE or Automatic selection (Area 2). The available resources on the heterogeneous ensemble are shown in Area 4.
- Select the neuronal simulation models and parameters (Area 3):
- The neuron model along with the model parameters.
- The plasticity mechanism, if any, or static synapses.
- The connections of the neuronal network and their parameters.
- The number of neuron cells and the simulation time of the network.
- Finally, run the simulation from the Actions buttons in Area 1, and download the results or reset the simulation options.
A DEMO video of the BrainFrame framework is available HERE
POSITIVE REACTION TO BRAINFRAME FROM STAKEHOLDERS The BrainFrame web service was presented by Dr. Christos Strydis at an annual lab talk of the Neuroscience Department in the Erasmus Medical Centre, on 18 of April 2017. The presentation addressed the benefits of heterogeneous computing for neuroscience and included a live demonstration of the web service to the audience. Various attendees requested for various simulation runs and were able to see the results being generated on the fly! Overall, the reception of BrainFrame was high and interest has peaked.
Cover page and sample page of lab talk |
Mr. Harry Sidiropoulos and Dr. Christos Strydis (Neurasmus)