guest ::
| Home > Publications database > NUMA balancing hampering performance of spiking network simulations |
| Home > Publications database > NUMA balancing hampering performance of spiking network simulations |
| Conference Presentation (After Call) | FZJ-2026-04464 |
2026
Abstract: Computing centers today mostly operate conventional CPU- and GPU-based systems, where thedirect way of reducing energy consumption is a reduction in the applications’ runtime. Neuromor-phic computing promises energy efficient architectures for artificial intelligence. In this endeavor,code for the simulation of large-scale spiking networks on conventional supercomputers is the ref-erence. We show that turning off automatic NUMA balancing may reduce energy consumption by20%. This dwarfs other attempts of cost-effective efficiency measures of a computing center. Thememory access pattern of spiking network simulation code dynamically interacts with automaticNUMA balancing. This does not affect the correctness of simulation results and thus goes unno-ticed in day-to-day neuroscience research. In performance analysis, however, time measurementsfluctuate, obstructing attempts to optimize simulation technology. A new time- and compute-noderesolved performance display exposes fine-grained temporal variability during distributed spikingnetwork simulations. The analysis shows that automatic NUMA balancing is disadvantageous, partic-ularly affecting the jemalloc library for thread-aware memory allocation in a transient manner. Thenew method also allows developers to detect perturbations of HPC systems and target improvementsto simulation technology. As a consequence, we have equipped our supercomputer JURECA with anoption to turn on or off automatic NUMA balancing per-node at user level. This gives researchers theopportunity to optimize settings for their applications. There are indications in the literature that theeffect has been observed before, yet it does not seem common knowledge in scientific computing. Itremains to be investigated how widespread the phenomenon is among scientific codes.
|
The record appears in these collections: |