LSPP 2015 - Workshop on Large-Scale Parallel Processing

While much attention has been focused on the grand challenge of scaling compute capability to Exascale systems, there has been less hoopla around how to scale the interconnection networks for such beasts. While historical supercomputer bytes-per-flop (or per-op) ratios remain desirable, there are a number of formidable challenges to this nirvana. At the top of the list, the cost per byte and energy per byte for node-to-node communication is not on the same steep curve as compute engines. Additionally, network hardware and software will be stressed by the growing number of network endpoints: 10s or 100s of millions of compute engines are required to reach Exascale. Reliability is yet another challenge that is growing with both system scale and semiconductor process scaling. In this talk we will outline a set of network ideas and directions to overcome or cope with these realities. Along the way, we will take a closer look at several areas which will impact Exascale network scaling: optics technologies, topologies, accelerators, and scalable messaging architecture.