Paper Title: Identifying Workloads Well-Suited to Acceleration by IPUs

Paper Abstract: IPUs can optimize total cost both by offloading common tasks and by using targeted acceleration to optimize the performance of key workloads. Choosing the best-suited workloads is a key step in broadening the scope and application of IPUs. The performance analysis of a collection of common workloads spanning networking, crypto, and storage illustrates ways to make the best selections. Example workloads include container networking, load balancers DPDK-based IO forwarding, and NVMe/TCP initiators. The results will help designers address the challenges in offloading onto IPUs and suggest the best benchmarks to use in characterizing IPU performance.

Paper Author: Yadong Li, SW Architect, Intel

Author Bio: Yadong Li is a Principal Engineer at Intel and the company’s lead software architect for next generation IPU architecture and design. His focus is on storage technologies and architecture. He led Intel’s Mount Evans (the company’s first ASIC-based IPU) software architecture from concept to product shipping and received an Intel Achievement Award for his work. He has a deep understanding of both software and hardware architecture, software/hardware co-design, and system level performance optimizations. He holds over 17 granted patents and over 20 pending ones. He earned an MS in Computer Science from Wright State University (Dayton, OH).