Thursday, June 15th
02:00-3:20 PM
C-202: Academic Research Review (Academic Track)
Paper Title: New TCP Stack Offloads Content Delivery I/O to a SmartNIC

Paper Abstract: Recent advances in high-bandwidth I/O devices enables scalable delivery of online content. However, most content servers run slowly because they overburden the CPU with disk and network I/O (over 70% of CPU cycles in our experiments). A new stack, IO-TCP, has a split design that reduces the burden on the CPU by offloading disk I/O and TCP packet transfer overhead to a SmartNIC. The result is to separate the task’s control and data planes to raise performance. An evaluation shows that TLS file transfer runs twice as fast on an I/O-TCP-ported lighttpd than on standard Linux TCP even on an Atlas server using 10 cores.

Paper Author: Taehyun Kim, Graduate Student, KAIST

Author Bio: Taehyun Kim is a doctoral student at Korea Advanced Institute of Science and Technology (KAIST, Daejeon, Korea). His research interests are efficient data center networking, the impact of operating system-level behavior in servers on network performance, and efficient, secure learning and inference frameworks for deep learning AI models. He has been an intern at Microsoft Research, where he studied the optimization of systems for training and serving deep learning models. He has published this paper at the USENIX Symposium on Networked Systems Design and Implementation (NSDI).