Workshop: Workshop on High Performance Fabrics for AI and HPC
Authors: Vijay Thurimella (HPE, Georgia Institute of Technology) and Gourav Rattihalli, Pavana Prakash, and Dejan Milojicic (HPE)
Abstract: Rack scale AI and HPC systems combine accelerators with high bandwidth fabrics but often impose fixed resource ratios and communication domains. We evaluate eleven applications and two bounds across training, inference, molecular dynamics, linear algebra, and graph analytics while varying host resources, accelerator count, and connectivity. Fabric sensitivity differs sharply. Disabling peer to peer communication reduces all to all bandwidth by 36.1× and all reduce by 16.3×; HPL fails to initialize and training slows 2.23×, while GPU HPCG and DeepDriveMD are unaffected. These results motivate software composition for resource changes and live optical reconfiguration for connectivity changes without destroying state. Our prototype matches native bandwidth, but state reconstruction adds substantial delay. Dynamic composition surpasses static goodput near 72 second idle windows and reaches 1.34× at 160 seconds. Reconfigurable fabrics therefore require topology, opportunity duration, predictability, and state preservation.
Back to Workshop on High Performance Fabrics for AI and HPC Archive Listing
Back to Full Workshop Archive Listing