Disabling NUMA balancing cuts energy 30% in spiking simulations

A team of 12 researchers, led by Melissa Lober at the Jülich Supercomputing Centre, has discovered a simple yet powerful energy-saving lever for high-performance computing workloads. In their paper on arXiv (2607.22275), they show that turning off automatic NUMA (Non-Uniform Memory Access) balancing reduces energy consumption by 30% for large-scale spiking neural network simulations. These simulations serve as the benchmark for neuromorphic computing, which promises energy-efficient AI hardware. The finding dwarfs other optimization attempts in cost-effectiveness—and the fix is just an OS setting. The team equipped their supercomputers with a per-job toggle, letting users choose the best setting for each application.

The problem arises because automatic NUMA balancing dynamically migrates memory pages to balance access times—but for spiking network codes, memory access patterns interact with this balancing in a way that degrades performance without affecting correctness. This went unnoticed in neuroscience research but caused large fluctuations in timing measurements, hampering optimization. The performance display revealed fine-grained temporal variability and showed that NUMA balancing transiently harms the jemalloc thread-aware memory allocator. The paper notes the effect may have been observed before, but it remains underappreciated in scientific computing. The implication: a straightforward OS-level tweak can slash HPC energy costs, with potential applicability to other scientific codes.