HyperX Network Study Unveils Diagonal Allocation Beating Traditional HPC Strategies

As high-performance computing (HPC) systems scale, minimizing communication overhead becomes critical. HyperX networks offer a richly connected, low-diameter topology that is both scalable and cost-effective compared to Torus, Fat-tree, or Dragonfly. However, resource allocation strategies for HyperX were previously underexplored, and methods from other topologies don't transfer directly. In this arXiv paper, Alejandro Cano and colleagues formalize allocation strategies for HyperX, categorizing them into linear, geometric, and stochastic functions. They characterize strategies through theoretical analysis of dilation, convexity, and partition properties, then evaluate them with synthetic traffic and real application kernels under various routing algorithms.

The study reveals that partition bandwidth and switch locality are decisive in mitigating interference. Notably, the Diagonal allocation strategy—which is not convex—consistently outperforms traditional convex approaches in most scenarios. The authors provide lessons learned for implementing resource allocation policies in HyperX-based HPC systems. This work fills a key gap in HPC network research, offering practical guidance for system architects and job schedulers aiming to reduce latency and improve throughput on HyperX topologies.