NEST performance benchmarks

NEST performance is continuously monitored and improved across various network sizes. Here we show benchmarking results for NEST version 3.8 on Jureca-DC [1]. The benchmarking framework and the structure of the graphs is described in [2].

Strong scaling experiment of the Microcircuit model [3]

_images/mc_benchmark.png

  • The model has ~80 000 neurons and ~300 million synapses, minimal delay 0.1 ms

  • 2 MPI processes per node, 64 threads per MPI process

  • Increasing number of computing resources decrease simulation time

  • Data averaged over 3 runs with different seeds, error bars indicate standard deviation

  • The model runs faster than real time [4]

Strong scaling experiment of the Multi-area-model [5]

Dynamical regime: Ground state

_images/mam_ground-state_benchmark.png

Dynamical regime: Metastable state

_images/mam_metastable-state_benchmark.png

  • The model has ~4.1 million neurons and ~24 billion synapses, minimal delay 0.1 ms

  • It can be run in two different dynamical regimes: the ground state and the metastable state [5].

  • 2 MPI processes per node, 64 threads per MPI process

  • Steady decrease of run time with additional compute resources

  • Data averaged over 3 runs with different seeds, error bars indicate standard deviation

Weak scaling experiment of the HPC benchmark model [6]

_images/hpc_benchmark.png

  • The size of network scales proportionally with the computational resources used

  • Largest network size in this diagram: ~5.8 million neurons and ~65 billion synapses, minimal delay 1.5 ms

  • 2 MPI processes per node, 64 threads per MPI process

  • The figure shows that NEST can handle massive networks and simulate them efficiently

  • Data averaged over 3 runs with different seeds, error bars indicate standard deviation

See also

Example networks:

References