Title: 

Building the Next Generation Internet Router: Practical GPU Based Packet Processing

 

Abstract: 

The telecom industry is transitioning from traditional network hardware to software defined networking on commodity hardware. Many key packet processing algorithms can be made highly parallel and have already been ported to leverage GPU parallel computing. Numerous studies show large gains in the price/performance ratio, but simplifying assumptions about the data path, often made when evaluating parallel packet processing algorithms on GPUs, cast doubt on the achievable benefits of parallel packet processing in a real world production environment. In order to facilitate evaluations of packet processing on GPUs, many studies isolate the computational components and ignore the time intensive process of transferring data between the NIC, CPU and GPU. Our research focuses on an end-to-end solution for GPU packet processing at line speeds above 10 Gbps that includes streaming data from the NIC directly into the GPU, processing the packets, and streaming the results back out onto the network. By analyzing and optimizing the entire packet processing pipeline, the telecommunications industry can reap the enormous price/performance gains promised by using GPUs to process network packets.

 

Bio:

Minsung Jang is a Principal-Inventive Scientist at the AT&T Labs - Research. His current research concerns software methods that accelerate the data plane in software-defined datacenters with the next-gen programmable devices including GPUs, FPGAs, and merchant silicon. He is also responsible for architecturing the edge cloud infrastructure for 5G services. Before joining AT&T, he worked at Samsung Electronics, IBM Research, Nokia Research Center, and Intel Labs. He obtained his Ph.D. in computer science at the Georgia Institute of Technology and his undergraduate and master’s degrees from Yonsei University, Seoul, Korea.