EDEN: A High-Performance, General-Purpose, NeuroML-Based Neural Simulator

Abstract

Modern neuroscience employs in silico experimentation on ever-increasing and more detailed neural networks. The high modeling detail goes hand in hand with the need for high model reproducibility, reusability and transparency. Besides, the size of the models and the long timescales under study mandate the use of a simulation system with high computational performance, so as to provide an acceptable time to result. In this work, we present EDEN (Extensible Dynamics Engine for Networks), a new general-purpose, NeuroML-based neural simulator that achieves both high model flexibility and high computational performance, through an innovative model-analysis and code-generation technique. The simulator runs NeuroML-v2 models directly, eliminating the need for users to learn yet another simulator-specific, model-specification language. EDEN's functional correctness and computational performance were assessed through NeuroML models available on the NeuroML-DB and Open Source Brain model repositories. In qualitative experiments, the results produced by EDEN were verified against the established NEURON simulator, for a wide range of models. At the same time, computational-performance benchmarks reveal that EDEN runs from one to nearly two orders-of-magnitude faster than NEURON on a typical desktop computer, and does so without additional effort from the user. Finally, and without added user effort, EDEN has been built from scratch to scale seamlessly over multiple CPUs and across computer clusters, when available.

Keywords: High-Performance Computing; NeuroML; biological neural networks; code morphing; computational neuroscience; interoperability; simulation; software.

Figure 1

A relative comparison of the characteristics of EDEN and the established neural simulators. (A) compares the simulators on the performance and usability plane; (B) shows the ordering between simulators regarding the level of model detail; and (C) shows the level of modeling detail supported by each simulator.

Figure 2

EDEN's processing pipeline. The whole model is analyzed in order to extract the computationally similar parts of neurons, and to generate optimized code and data representations for them, on the fly.

Figure 3

Code and data signatures for an exponential-conductance post-synaptic component (A), and for a classic Hodgkin-Huxley sodium channel (B).

Figure 4

The stages of the per-neuron signature synthesis process, for neurons with few (phenomenological) compartments. The neuron shown consists of three different compartments, each containing different physiological mechanisms. The simulation code for all mechanisms is laid out in a flat format, along with their associated data. Thus, a streamlined and compact code kernel is created for this specific type of neuron.

Figure 5

The stages of the per-compartment signature de-duplication process, from the abstract model to the concrete implementation. On the schematic of the detailed neuron model, distinct compartment types are shown in different colors. The components sharing the same type are then grouped together, in terms of simulation code and data representation. The specific mechanisms comprising the compartments and the data cells they contain are not shown here, for brevity.

Figure 6

A schematic representation of how the extracted work item signatures are converted to low-level data structures for efficient processing. For each work item, the set of scalars and work tables of each is appended into flat node-wide arrays, for each data type. Data types shown: CF32 = 32-bit floating-point constants, CF64 = 64-bit integer constants, SI64 = 64-bit integer variables. The different data types for scalars and tables have been omitted for clarity in the diagram, without loss of generality. Colors indicate different code-data signatures among work items.

Figure 7

Histograms of relative error under the NeuroML-DB similarity (A) and inter-spike interval (B) accuracy metrics for each NeuroML-DB neuron model. The bins around the “ < -10” and “>10” limits include all models with more than 10% of discrepancy.

Figure 8

Raster plots for each network used in the performance benchmarks, when run on NEURON and EDEN. Note that the input stimuli are pseudo-randomly generated.

Figure 9

Histograms of neural activity metrics for each network used in the performance benchmarks. FR, firing rate; LV, local variation; ISI, inter-spike interval; CC, short-time firing correlation; RC, rate correlation; (λ), correlation eigenvalue. For each simulated network, the solid green line outlines the distribution of metric values when using EDEN, and the overlaid dashed red line outlines the distribution when using NEURON. On the right, the effect size (ES) and its confidence interval is shown for each applicable metric.

Figure 10

Run time for each neural network considered, for NEURON versus jNeuroML/NEURON and versus EDEN on one CPU thread and on all CPU threads. For each neural network, benchmark the bar height in the chart is normalized against NEURON's run time for that benchmark.