Dynamic load balancing in parallel applications traditionally minimizes the maximum workload across processors. But in communication-intensive codes — particle-in-cell simulations, finite element methods — balancing compute load can increase communication overhead by scattering coupled objects across nodes.
The authors (arXiv:2603.23329) present a diffusion-based load balancer that simultaneously reduces communication and distributes work. The key: using the application's communication graph to guide redistribution. Objects that frequently exchange data are kept on the same node when possible. Load imbalance is corrected by diffusing excess work to neighboring nodes in the communication topology, not to random nodes.
The through-claim: optimal load balancing is not minimum makespan — it's minimum makespan plus communication cost. The conventional formulation treats computation and communication as independent objectives, but they're coupled through object placement. A processor that's slightly overloaded but communicates locally outperforms one that's perfectly balanced but communicates remotely. Diffusion respects this coupling because it moves work along communication edges, not across them.