FLEET—Fast Lanes for Expedited Execution at 10 Terabits: Program Overview

The DARPA FastNICs program targets orders of magnitude improvement in applications such as deep learning training by making radical improvements to network performance: while raw bandwidth has grown dramatically, the fundamental roadblock to application performance has been in delivering that data to the application. FLEET provides a primarily off-the-shelf solution with high-end servers and shared computational and storage resources connected via PCIe over a reconfigurable MEMS optical switch; it uses custom Optical NICs to allow arbitrary topologies that can be configured before or even during execution to take advantage of shared resources and to flow data between components. FLEET's software is derived from Stanford Legion, which we are modifying to use the FLEET hardware and to plan application execution for these dynamic network topologies.