SDI Seminar - Josh Fried

— 1:00pm

Location:
In Person - Panther Hollow Room 4105, Mehrabian Collaborative Innovation Center

Speaker:
JOSH FRIED, Ph.D. Student, Computer Science & Artificial Intelligence Laboratory, Massachusetts Institute of Technology
https://joshfried.io/

Making Kernel Bypass Practical for the Cloud with Junction

Kernel bypass systems deliver significant improvements in throughput and tail latency for network-intensive applications, often outperforming traditional operating systems (OSes) by an order of magnitude. However, these performance gains come at a steep cost: they rely on dedicated resources (e.g., spinning cores, pinned memory) and demand extensive application rewriting. These limitations make kernel bypass systems impractical for cloud operators, who prioritize dense application packing and seek to avoid costly code modifications. 

In this talk, I will introduce Junction, a kernel bypass system that overcomes these challenges, supporting thousands of instances on a single machine while remaining compatible with unmodified Linux applications. Junction leverages modern NIC features to minimize memory and monitoring overheads, enabling dense packing of instances. It incorporates a library operating system that implements the Linux system call interface, allowing compatibility with existing applications while relying on kernel bypass for efficient OS functionality. Junction matches or exceeds the performance of state-of-the-art kernel bypass systems without requiring application porting. It improves throughput, latency, and CPU efficiency for unmodified network-intensive datacenter applications and seamlessly supports popular frameworks like Go, Python, and Java. 

— 

Josh Fried is a senior PhD student at MIT CSAIL advised by Adam Belay. He is broadly interested in operating systems, networks, and distributed systems, with a focus on improving performance and efficiency for datacenters through better operating system design. 

Faculty Host:  Justine Sherry

Event Website:
https://www.pdl.cmu.edu/SDI/2024/120524.html