Emerging data center network architectures are becoming reconfigurable: Optical switches enable the network’s topology to be adjusted to accommodate skewed traffic demands or to recover from failures. While prior work has shown the practical benefits of reconfigurable topologies, several performance aspects are not yet well understood. In particular, most reconfigurable data center topologies are hybrid in nature, where parts of the network are reconfigurable (consisting of optical or wireless devices) while other parts are static (consisting of electrical switches). Current proposals treat each part separately and route flows on either part exclusively. We argue that such segregation in topology results in non-optimal paths for flows in the network. Moreover, multi-hop routing across optical switches in data center networks is currently relegated to either non-optimal schemes or intractable integer programs, even though having more than one optical switch is essential at scale. To solve the above two issues, we present the first algorithmic study of reconfigurable network architectures. We pinpoint the intractability barriers between static and reconfigurable parts of the network, complementing them with two optimal algorithms in terms of demand-awareness and average path lengths. More specifically, our results show that when the hybrid network parts are considered segregated, as in prior work, classic matching algorithms used in prior work are optimal if there is only one optical switch in the topology. However, when the hybrid network is seen from a joint perspective, optimal routing becomes an NP-hard problem. We further investigate networks with multiple optical switches. We find that finding optimal routes for even two flows in such networks is an NP-hard problem, but a single flow can be optimally routed.

Top