Abstract
Large-scale data centers for commercial workloads or high-performance computing can reap significant benefits by applying "slow" (ms-scale) optical circuit switching (OCS) to enable reconfigurable system partitioning. Presently available commercial OCS offerings with hundreds of ports can enable partitioning of systems with tens of thousands of endpoints. We argue for the usefulness of such partitioning, and demonstrate how to apply it to large-scale dragonfly networks. To make this approach practically viable for exascale systems, cost-effective OCS and single-mode WDM transceivers will be required.