Wait, how many networks did I just cross? Interconnects, Accelerators, and Disaggregation at TACC

Panelist – Dan Stanzione, Ph.D., Executive Director, Texas Advanced Computing Center

In the HPC world, there is an awful lot of demand for the PCI slots in our servers.  Accelerators, high-speed networks, and now PCI disaggregation, soon CXL, etc.  Often, these options can even be redundant (should you use your IB or switched-PCI connection for your MPI messages? How about I/O?).  This talk looks at the experiences at the Texas Advanced Computing Center in both older and newer high-speed networks (Infiniband, Omnipath, Rockport), how our job mix is evolving the traffic patterns, particularly with the rise of AI, and how we see our more recent experiments with node disaggregation over PCI fitting in.