Exploiting network topology in brain-scale simulations of spiking neural networks

A team of researchers including Melissa Lober, Markus Diesmann, and Susanne Kunkel has published a groundbreaking paper on arXiv that redefines the bottleneck in simulating brain-scale spiking neural networks. The work, titled 'Exploiting network topology in brain-scale simulations of spiking neural networks,' challenges the conventional wisdom that simulation speed is limited by supercomputer interconnect hardware or communication libraries. Through profiling, the team discovered the real culprit is the variability in computation time between nodes, leading to extensive waiting for the slowest participant. This synchronization overhead, inherent in standard collective communication calls, was the primary drag on performance.

The researchers' fundamental cure is to reduce the need for synchronization by cleverly mapping the simulation's structure to the computer's architecture. They observed that the mammalian brain is organized into areas with short internal delays and longer delays between areas. Their proposed 'structure-aware mapping' assigns entire brain areas to individual compute nodes, allowing for frequent local communication within a node and much less frequent global communication between nodes. This local-global hybrid approach led to a 'substantial performance gain' in a real-world test case. The work provides concrete guidelines for more energy-efficient simulations on conventional supercomputers and simultaneously raises the performance bar that emerging neuromorphic computing systems must meet.