Networking (Research activity)
Networking is probably one of the biggest challenge of the IT industry. With datacenter fast growth, and power centric solution, the number of nodes to interconnect moved from a few thousands to hundred of thousands during the past decade. Compute node size has decreased with higher integration of compute core per CPU, virtualization is stressing the infrastructure, and made growth the number of IP addresses to an unsustainable value within a single datacenter.
Traditionnal network topology is driven by system integration where a basic element is a server, standing into a rack equipped with a “natural” top of the rack switch which interconnect racks rows all together with an infrastructure rack, which goes to a back bone to get access to external ressources. This approach is mostly materialized by a network with a tree topology.
Supercomputers have introduced new topologies which improved latency, and bandwidth by using advanced protocols, and expanding Tree to Fat Tree (uplink bandwidths are the same than downlink), and multiplying network cables, which unfortunatly increased link failure rate, but without service interruption due to cabling redundancy. Cray and IBM worked on new system interconnect based on Torus topology which have the benefit to remove single point of failure from the network and simplify the switching hardware. Unfortunatly this topology doesn’t fit really well into racks and traditionnal datacenter.
RuggedPOD as being a new system form factor might be really suitable to 2D torus and 3D torus topologies. We are exploring this option into this page as to better understand and share with you the tradeoff behind this technology.