P076 A NEST-based framework for the parallel simulation of networks of compartmental models with customizable subcellular dynamics
Leander Ewert1, Christophe Blaszyck2, Jakob Jordan5, Charl Linssen1,3, Pooja Babu1,3, Abigail Morrison1,2, Willem A.M. Wybo4
1Institute for Advanced Simulation (IAS-6) and JARA-BRAIN Institute I, Jülich Research Centre, 52425 Jülich, Germany
2Department of Computer Science 3 - Software Engineering, RWTH Aachen University, Aachen, Germany
3Simulation and Data Laboratory Neuroscience, Jülich Supercomputer Centre, Institute for Advanced Simulation, Jülich Research Centre, 52425 Jülich, Germany
4Peter Grünberg Institut (PGI-15), Jülich Research Centre, 52425 Jülich, Germany
5Department of Physiology, University of Bern, Bern, Switzerland
*Email: l.ewert@fz-juelich.de
Introduction
The brain is a massively parallel computer. In the human brain, 86 billion neurons convert synaptic inputs into action potential (AP) output. Moreover, even at the subcellular level, computations proceed in a massively parallel fashion. Approximately 7’000 synapse per neuron are supported by complex signaling networks within dendritic compartments. In itself, these signaling networks can also be understood as nanoscale computers that convert synaptic input, backpropagating APs, and local voltage and concentration signals into weight dynamics that support learning and memory. It is only natural, thus, to use the parallelization and vectorization capabilities of modern supercomputers to simulate the brain in a massively parallel fashion.
Methods
The NEural Simulation Tool (NEST) [1] is the reference with regards to the massively parallel simulation of spiking network models, as it has been optimized to efficiently communicate spikes across MPI processes [2]. Moreover, these capabilities introduce little overhead for the user, as the distribution of neurons across MPI processes is taken care of by NEST itself. However, so far NEST had limited options to simulate subcellular processes as part of the network, essentially forcing users to develop custom C++ codes. We have extended the scope of the NESTML modelling language [3] to support multi-compartment models, with dendrites featuring user-specified dynamical processes (Fig 1A-C).
Results
These user-specified dynamics are compiled into efficient NEST models through a C++ code generator, in such a way that the vectorization capabilities of modern CPUs are optimally leveraged. This allows for a deeper level of parallelization, next to the network parallelization across MPI processes, allowing individual CPUs to integrate up to eight compartments in parallel and decreasing single neuron runtimes accordingly. The compartmental engine furthermore leverages the Hines algorithm [4] to achieve stable and efficient integration of the system as a whole. Together, this results in single-neuron speedups compared to the field-standard NEURON simulator [5] of up to a factor of four to five (Fig 1D).
Discussion
Thus, we enable the simulation of large-scale networks where individual neurons have user-specified dynamical processes, representing (i) voltage-dependent ion channels, (ii) synaptic receptors that may be subject to a-priori arbitrary plasticity processes, or (iii) slow processes describing molecular signaling or ion concentration dynamics. Conducting such simulations has historically been challenging, since simulators specific to this purpose were lacking. With the present work, we facilitate the creation and efficient distributed simulation of such networks, thus supporting the investigation of the role of dendritic processes in network-level computations involving learning and memory.
Figure 1. Figure 1. (A) NESTML-defined subcellular mechanisms (left) are compiled into an efficient NEST model. User-defined dendritic layouts (middle) are then embedded in NEST network simulations (right). (B) NESTML code defining dendritic calcium dynamics induces BAC firing [6] in a two-compartment model [7] (C). (D) Speedup of NEST compared to NEURON (bottom) for two dendritic layouts (left vs right).
Acknowledgements
The authors gratefully acknowledge funding from the HelmHoltz POF IV, Program 2 Topic 3.
References
[1] 10.4249/scholarpedia.1430
[2] 10.3389/fninf.2014.00078
[3] 10.3389/fninf.2018.00050
[4] 10.1017/CBO9780511541612
[5] 10.1016/0020-7101(84)90008-4
[6] 10.1038/18686
[7] 10.48550/arXiv.2311.0607
