Over the past 30+ years, I/O workflows have evolved and changed dramatically.  Data access via file protocols has morphed to Data access via object protocols. Permission granularity control requirements have grown from modebits to Access Control Lists to Identity Policies.  The sheer size of the datasets both in capacity and number of items has grown exponentially. Modern workflows, including those related to AI and Machine learning, involve access to data across multiple protocols with vastly different methods of permissioning and control.

This presentation walks through how data flows through GPUs during Large Language Model training. There is an introduction to LLM neural networks and how they map to the GPU arrays. We look at the challenges of current GPU topologies and look at how these topologies will evolve with the introduction of UEC and UALink.

Accessing servers from Linux over SMB3.1.1 continues to improve in exciting ways. This talk will explore the latest enhancements to the Linux SMB3.1.1 client, enabling more secure, efficient, and compatible access to remote storage across a wide range of SMB3 file servers—including Samba, Azure (the "largest server in the world"), ksmbd, Windows, NetApp, and macOS. There are also exciting improvements to smbdirect (SMB over RDMA), including making this high-performance path easier to use for both the client and server, as well as userspace applications like Samba.