Paper Title: Offload More Functions to DPUs Using On-Board Low-Power Cores

Paper Abstract: DPUs provide an efficient way to reduce virtualization overhead by implementing key functions with programmable cores and accelerators. Traditionally, the functions have been networking tasks such as virtualizing IO, switching, and overlays. However, other functions that typically run on the host can be offloaded to efficient general purpose cores, including storage stacks, security, API gateways, load balancers, and content managers. The cores must implement such higher layer functions at a reasonable value of performance/watt. They must also support a wide ecosystem of operating systems, packet processing frameworks, algorithms/libraries/tools, network/storage stacks, and applications that make deploying such functions on a DPU convenient and quick. Developers must identify the cores best suited to their application needs.

Paper Author: Matthew Dirba, Staff Design Engineer, Arm

Author Bio: Matthew Dirba is a Staff Design Engineer at ARM, where he works on performance validation, workload characterization, and methods for estimating and projecting performance-power ratios. He also has experience with developing firmware, drivers, and application software for server and embedded applications, including real-time. He earned a BS in computer engineering from Texas A&M University. He holds a patent for dynamic loading of software and has several awards and certifications from IBM.