Accelerating Data Serialization and Deserialization Protocols with In-Network Compute

Explore a 24-minute conference talk presented at the ExaMPI workshop during SC'22, focusing on accelerating data serialization and deserialization protocols using in-network compute. Dive into the research conducted by Shiyi Cao, Salvatore Di Girolamo, and Torsten Hoefler from the Scalable Parallel Computing Lab at ETH Zurich. Learn about the challenges of efficient data communication in datacenter and HPC systems, and discover how offloading data deserialization to programmable Smart-NICs can significantly improve performance. Examine the implementation of Google Protocol Buffers on Smart-NICs and understand how this approach achieves up to 4.8x higher throughput compared to a single AMD Ryzen 7 CPU. Gain insights into microservice throughput modeling and the benefits of pipelining deserialization with application activities using PsPIN. Follow the presentation's structure, covering topics such as data SerDes in distributed applications, performance bottlenecks, network acceleration, offloading overview, and evaluation results.

Intro
Data SerDes in distributed applications
Performance Bottleneck
Inefficient processing in general-purpose CPUs
Network Acceleration: SmartNIC Specialties
Offloading Overview
Decoding on the NIC (some examples)
Evaluation
Microservice E2E Throughput Modeling
Conclusion