I think the post is right - we need to be rethinking some of the abstractions in data center networking so we can deploy apps with specific network topologies and performance guarantees, and still be logically isolated from each other. This is especially important for bandwidth and latency sensitive apps like scientific computing that really care about what's happening at the "Top of the Rack" switches. I'm not sure that VLANs alone cut it.