Scalable and low server-to-server latency data center network architecture based on optical packet inter-rack and intra-rack switching

Abstract

In this work, we introduce an efficient data center network (DCN) architecture using optical packet switching for the inter-rack and intra-rack packet networks. We investigate the end-to-end communication in a server-to-server (S2S) base, implementing the east-west networking scenario across the whole intra- and inter-rack DCN. As opposed to other optical or hybrid optical–electrical DCN studies that focus on either the intra-rack or inter-rack part of the DCN, our study proposes and investigates a unified DCN architecture that consists of three separate optical network models: the intra-rack, the inter-rack, and the bridge that connects the intra- and inter-rack networks. Particularly, the intra-rack optical network is a passive-coupler-based single-hop wavelength division multiplexing (WDM) network for the communication among servers of the same rack, following bandwidth-efficient synchronous transmission WDM access (WDMA) and time division multiplexing access (TDMA) rules. The bridge optical network is designed as a passive-optical-network-based network to connect the rack servers with the above-placed top-of-rack (ToR) switch, bridging the intra- and inter-rack optical networks and following a greedy TDMA scheme. Finally, the inter-rack optical network connects the different ToRs in a 2D torus topology over optical fibers, offering all-to-all connectivity via lightpaths that rely on a combination of spatial and wavelength paths. In our study, the DCN traffic is classified into several priority classes, each representing distinct service delay requirements, as occurs in existing DCNs. The DCN architecture design, i.e., the server and ToR switch architectures as well as the TDMA/WDMA algorithms for the intra-rack, inter-rack, and bridge optical networks, takes into consideration the traffic variability aiming to serve it into a considerably low end-to-end latency time of the order of few µs, even under high congestion conditions. The proposed DCN performance is evaluated under the scenario of 400 Gbps and 8 Tbps total capacity in the intra-rack and whole end-to-end networks, respectively, while its limitations are extensively explored. Simulation results demonstrate that our proposal achieves 90% and 100% bandwidth utilization in the optical intra- and inter-rack networks, respectively, and 91% for end-to-end S2S communication across the whole DCN. Also, the maximum end-to-end packet latency experienced across the whole DCN under highly loaded conditions is only 0.98 µs, 27 µs, and 218 µs for the highest, medium, and lower priority traffic classes, respectively, fully complying with the rigid latency requirements of various modern cloud applications such as Industry 4.0 and the Internet of Things. The proposed DCN architecture is scalable and can accommodate more than 10,000 servers. In addition, it provides a low energy footprint ensuring up to 50% power consumption reduction as compared to existing Fat-Tree DCN architectures. Finally, it provides lower end-to-end latency across the whole DCN up to high loads, as compared with other relative studies.