DCell: A Scalable and Fault-Tolerant Network Structure for Data Centers

This paper attempts to solve the same problem as A Scalable, Commodity Data Center Network Architecture: the high cost of high-end non-commodity switches and the lack of full bandwith across the entire datacenter. DCell takes the network architecture into the servers. DCell also requires that each server have 2 GigE connections, but this is achieved with a readily available quad port Ethernet adapter. In data centers with tens to hundreds of thousands of servers, it would be desirable to cost-reduce with a Dual Port Ethernet adapter. The modifications to servers would not likely be a drawback since such a setup would be used in a new deployment (the interconnection costs have already been sunk in existing deployments).

The authors focus on three design goals for data center networks (DCN). First, the network infrastructure must be scalable to a large number of servers and allow for incremental expansion. Second, DCN must be fault tolerant against various types of server failures, link outages, or server-rack failures. Third, DCN must be able to provide high network capacity.

The main benefits of the DCell approach is that it has a great degree of fault tolerance and can scale to millions of servers with only a level 3 DCell and truly commodity 8-port switches.

The main drawback of this approach is the complexity of the wiring interconnections for even a DCell structure, especially when repeated many times, and up to a level 2 or higher DCell. This provides an opportunity for modular data centers pre-configured in DCells from the manufacturer. These modular DCells would have clear instructions (possibly color-coded) for interconnecting the modular DCells to one another and to the outside world.