Optical network topology design to execute many tasks simultaneously in a disaggregated data center

Abstract

In a disaggregated data center (DDC), task execution is reliant on the communication between resources, making performance highly sensitive to network quality. An optimized physical network topology is crucial for a DDC. To enable the simultaneous execution of numerous tasks, a substantial number of communicable resource pairs satisfying performance requirements is necessary. We propose a physical topology evaluation metric called the capability of simultaneous task execution (CSTE) and a corresponding physical topology design leveraging CSTE for a DDC equipped with optical networks. CSTE represents the ratio of resources that could be used as a resource communicating with other resources without violating the performance requirements in a situation where tasks up to the maximum number of executable tasks are executed. In addition, we formulated a physical topology design problem aimed at generating a physical network topology capable of maximizing task execution based on CSTE. By solving this optimization problem, we generated topologies and validated their effectiveness via task allocation simulations. The results showed that an optimal topology based on CSTE reduces task blockages by over 50% compared to conventional topologies. In addition, the results exhibited a positive correlation with the number of executable tasks. Through a physical topology design based on CSTE, we could construct a DDC that could handle a larger volume of tasks.