Network Performance Trade-Off in Modular Data Centers With Optical Spatial Division Multiplexing

Abstract

Modular design based on spatial division multiplexing switches is a promising way to improve the capacity and reduce the cabling complexity of data center networks. However, due to the coexistence of mice and elephant flows, in modular data center networks a trade-off between the blocking probability and total throughput arises. In fact, blocking elephant flows would lead to a relevant penalty on the throughput but a small penalty on the blocking probability. In this paper, we investigate the relation between the blocking and throughput in modular data center networks based on optical spatial division multiplexing. We combine the two metrics linearly by a weight factor that prioritizes them relatively. To solve the resource allocation problem, we propose both mixed integer linear programming formulations and close-to-optimal heuristics for three different spatial division multiplexing switching schemes. Simulation results demonstrate that a carefully chosen weight factor is necessary to achieve a proper balance between the blocking probability and throughput for all the schemes.