VL2: A Scalable and Flexible Data Center Network

[paper]

A basic tree topology for a data center partitions intra-DC communication. It's slow for the "far left" server to talk to the "far right" server. The authors of this paper want a uniform communication cost between all hosts in the database to promote agility, which means that any host can run any service at any given time. (This means that a DC can dynamically handle spikes in loads; no pre-allocation is necessary.) For this end, they suggest use of the Clos network topology, which essentially means that each aggregation switch is connected to all of the (many) intermediate switches. Shortest path first routing algorithms are inadequate here; they use Valiant Load Balancing (VLB) for routing so that traffic can be routed non-uniformly. VLB basically chooses a random path from the set of equivalent paths for a flow so that no particular path becomes overwhelmed. They call the use of VLB over a Clos network "VL2".

They do some measurements on an unspecified Microsoft data center to motivate their work. They show that 80% of DC traffic is internal, meaning you do want to focus on decreasing intra-DC transit costs. They also show that flows are small and uniformly distributed, meaning that (A) the system can work by load balancing flows instead of individual packets and (b) you don't need to support locality for the intra-DC traffic. They did a traffic analysis and found no regularity or patterns in the traffic, and they also found that traffic needs shifted a lot. I think this is a really good approach to motivating research -- they have actual hard evidence that the problem they are trying to solve exists. (When reading papers outside of my own area, I am nervous about accepting whether a problem really exists -- since I have no knowledge to judge it myself.) The one downside is that they don't say WHAT application this is for, so maybe you would see totally different patterns with a different type of application.