Singularity of Future Computer-System Networks

Abstract

As supercomputers and high-end datacenters become large, e.g. over million cores using over 20 millionWatts, their network design is more complex. In this talk, I will review their network requirements and trends, especially network topology for low communication latency and low-power consumption. Interestingly, these requirement may not accept traditional network topologies, such as k-ary n-cubes. In addition, for a diverse application workload, there are downsides to having a single network topology on a supercomputer. These facts give us a chance to innovate new computer-system networks that are quite different from existing supercomputer and datacenter networks. I recommend three unique technologies recently developed or reborn by my research group: (1) random network topologies, (2) free-space optics (FSO) and (3) in-water computers for making future supercomputers and datacenters. • Random network topologies: To achieve low latency, e.g. one microsecond for 100,000 endpoints, a topology of switches should thus have low diameter and low average shortest path length, both measured in numbers of switch hops. This can be modeled as an open problem to find a graph that has smallest diameter and average shortest path length given an order and a degree. The solution could be a non-structured (random) graph that will bring unique supercomputer and datacenter network topologies. • Free-space Optics (FSO): Using Free Space Optical (FSO) technology can replace a fraction of the inter-cabinet optical cables by FSO links. The endpoints of these FSO links can be swapped dynamically to transform the topology into a network topology optimized to a target parallel application. Our FSO prototype links use commodity optical transceivers and can achieve error-free communication for distances up to about 50 meters. In addition, the advantage of FSO links can be discussed in terms of reduced aggregate cable length. •In-water Computers and Networks: The computer cooling will become one of main concerns, because supercomputers and datacenters scaling is limited by the power budget. Surprisingly, more than 40% of total power consumption is sometimes consumed by cooling purpose. In this context, submerged computers are theoretically a good choice, because they make the best use of cold natural water. We have developed a prototype of submerged computers using a special resin coating. It currently works under the tap water.