A Lightspeed Data Center Network

Emerging data center applications demand low latency and high bandwidth
networks - similar to those found in high-performance computers.  This talk walks through a thought experiment of what a data center network using best practices HPC network design would look like. It shows that a “dragonfly network” using global adaptive routing and speculative reservations for congestion avoidance can offer network latencies that are dominated by the time-of-flight over the network cables.  Because of reduced buffering and better channel utilization such a network would have lower component cost than a conventional network with comparable performance.