A Multi-Stage Packet-Switch Based on NoC Fabrics for Data Center Networks

Bandwidth-hungry applications such as Cloud computing, video sharing and social networking drive the creation of more powerful Data Centers (DCs) to manage the large amount of packetized traffic. Data center network (DCN) topologies rely on thousands of servers that exchange data via the switching backbone. Cluster switches and routers are employed to provide interconnectivity between elements of the same DC and inter DCs and must be able to handle the continuously variable loads. Hence, robust and scalable switching modules are needed. Conventional DCN switches adopt crossbars or/and blocks of memories in multistage interconnection architectures (commonly 2-Tiers or 3-Tiers). However, current multistage packet switch architectures, with their space-memory variants, are either too complex to implement, have poor performance, or not cost effective. In this paper, we propose a novel and highly scalable multistage packet-switch design based on Networks-on-Chip (NoC) fabrics for DCNs. In particular, we describe a novel three-stage packet-switch fabric with a Round-Robin packets dispatching scheme where each central stage module is based on a Unidirectional NoC (UDN), instead of a conventional single hop crossbar fabric. The proposed design, referred to as Clos- UDN, overcomes all the shortcomings of conventional multistage architectures. In particular, as we shall demonstrate, the proposed Clos-UDN architecture: (i) Obviates the need for a complex and costly input modules, by means of few, yet simple, input FIFO queues. (ii) Avoids the need for a complex and synchronized scheduling process over a high number of input-output modules and/or port pairs. (iii) Provides speedup, load balancing and path-diversity thanks to a dynamic dispatching scheme as well as the NoC based fabric nature. Extensive simulation studies are conducted to compare the proposed Clos-UDN switch to conventional multistage switches. Simulation results show that the Clos-UDN outperforms conventional design under a wide range of input traffic scenarios, making it highly appealing for ultra-high capacity DC networks