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-2026-04463 |
2026
Abstract: Computing centers today mostly operate conventional CPU- and GPU-based systems, where the di-rect way of reducing energy consumption is a reduction in the applications’ runtime. Neuromorphiccomputing promises an alternative architecture with improved energy efficiency for artificial intel-ligence. In this endeavor, code for the simulation of large-scale spiking networks on conventionalsupercomputers is the reference. We show that turning off automatic NUMA balancing may reduceenergy consumption by 20%. This dwarfs other attempts of increasing the energy efficiency of acomputing center with respect to cost effectiveness.The typical memory access pattern of spiking network simulation code dynamically interacts with au-tomatic NUMA balancing. This does not affect the correctness of simulation results and thus goes un-noticed in day-to-day neuroscience research. In performance analysis, however, time measurementsfluctuate obstructing attempts to optimize simulation technology. A new time- and compute-node re-solved performance display exposes the fine-grained temporal variability in the course of distributedspiking network simulations. The analysis uncovers that automatic NUMA balancing is of disadvan-tage and, in particular, affects the jemalloc library for thread-aware memory allocation in a transientmanner. The new method also allows developers to detect perturbations of the HPC system and targetspecific improvements to simulation technology.As a consequence of these findings we have equipped our supercomputer JURECA with an optionto turn on or off automatic NUMA balancing on a per-node basis on the user level. This gives re-searchers the opportunity to find the best setting for the application at hand. There are indications inthe literature that the effect has been observed before, yet it does not seem common knowledge in sci-entific computing. It remains to be investigated how widespread the phenomenon is among scientificcodes
|
The record appears in these collections: |