Presentation
(PhD09) Bringing Neuroscience to HPC
SessionPhD Forum
Event Type
PhD Forum
Brain Simulations
Containerized HPC
HPC Accelerators
Heterogeneous Systems
Performance Analysis and Optimization
TimeMonday, June 25th1:33pm - 1:37pm
LocationAnalog 1, 2
DescriptionThe poster aims at presenting a collaborative effort between the National Technical University of Athens (NTUA) and the Erasmus Medical Centre Rotterdam (EMC Rotterdam) to develop a platform for the neuroscientific community that allows conducting of computationally-heavy experiments regarding the functionality of the human brain.
The author of the poster and NTUA's Ph.D. student, George Chatzikonstantis, develops accelerated backend solvers for typical workloads in the domain of computational neuroscience, which aim in particular at tackling biophysically accurate and complicated neuron models at a scale previously thought unapproachable. With the help of modern HPC fabrics, such as Intel's Xeon Phi manycore (co-)processors, network solvers with impressive speed are attainable, allowing carrying out neuroscientific experiments of meaningful accuracy and massive scale, offering a solid alternative to usual black-box approaches. Our efforts so far show simulations of networks scaling up to 1 million neurons and 1 billion synapses as possible even on a single accelerator.
Stemming from the collaboration between the author and other developers of neuroscientific solutions on HPC fabrics (chiefly from Erasmus MC Rotterdam), a more complete solution in the domain of computational neuroscience is being developed. BrainFrame is an online system which utilizes a heterogeneous cloud of backends, with simulators running on GPUs, FPGAs, Xeon Phis and even traditional high-end CPUs. The computational power of the heterogeneous cloud allows BrainFrame to provide the correct computing fabric depending on the experiment the user desires to submit to the online platform. In our studies, we stress the importance of utilizing the correct backend technology depending on the characteristics of each experiment (network size, connectivity density, neuron model level of detail).
The entire system is being developed in a dockerized environment in order to be made modular; the goal of BrainFrame is to continuously integrate new neuron models and corresponding backend solvers. We envision a potential heterogeneous system which allows the simulation of any neuronal model in massive quantities, without forcing the neuroscientist (nor his lab) to spend time and money in difficult-to-obtain equipment, hardware and coding experience.
The BrainFrame system is currently under development. A proof of concept has been already published in the Journal of Neural Engineering 2017 titled "BrainFrame: a node-level heterogeneous accelerator platform for neuron simulations". The aforementioned partners are currently working at releasing an alpha version where neuronal network simulators will be available and easily usable online through our hardware, as well as providing accelerated simulator alternatives (featuring orders of magnitude improvements in performance) for a subset of the supported neuron models.
The author of the poster and NTUA's Ph.D. student, George Chatzikonstantis, develops accelerated backend solvers for typical workloads in the domain of computational neuroscience, which aim in particular at tackling biophysically accurate and complicated neuron models at a scale previously thought unapproachable. With the help of modern HPC fabrics, such as Intel's Xeon Phi manycore (co-)processors, network solvers with impressive speed are attainable, allowing carrying out neuroscientific experiments of meaningful accuracy and massive scale, offering a solid alternative to usual black-box approaches. Our efforts so far show simulations of networks scaling up to 1 million neurons and 1 billion synapses as possible even on a single accelerator.
Stemming from the collaboration between the author and other developers of neuroscientific solutions on HPC fabrics (chiefly from Erasmus MC Rotterdam), a more complete solution in the domain of computational neuroscience is being developed. BrainFrame is an online system which utilizes a heterogeneous cloud of backends, with simulators running on GPUs, FPGAs, Xeon Phis and even traditional high-end CPUs. The computational power of the heterogeneous cloud allows BrainFrame to provide the correct computing fabric depending on the experiment the user desires to submit to the online platform. In our studies, we stress the importance of utilizing the correct backend technology depending on the characteristics of each experiment (network size, connectivity density, neuron model level of detail).
The entire system is being developed in a dockerized environment in order to be made modular; the goal of BrainFrame is to continuously integrate new neuron models and corresponding backend solvers. We envision a potential heterogeneous system which allows the simulation of any neuronal model in massive quantities, without forcing the neuroscientist (nor his lab) to spend time and money in difficult-to-obtain equipment, hardware and coding experience.
The BrainFrame system is currently under development. A proof of concept has been already published in the Journal of Neural Engineering 2017 titled "BrainFrame: a node-level heterogeneous accelerator platform for neuron simulations". The aforementioned partners are currently working at releasing an alpha version where neuronal network simulators will be available and easily usable online through our hardware, as well as providing accelerated simulator alternatives (featuring orders of magnitude improvements in performance) for a subset of the supported neuron models.
Poster PDF
Authors
