On the performance of scalable fast optical switch based recursive network architecture for HPC applications

We investigate a novel scalable low-latency HPC network for blade-to-blade interconnection exploiting multiple-parallel recursive architecture with fast optical switches under realistic HPC application traffics. Compared with Leaf-Spine HPC interconnecting 4096-blade, results show that HFOS4 reduces up to 75.9% latency, 64.2% cost and 30.9% power consumption.