CSE Doctoral Student Seminar: Adam Drescher and Hang Yan

Nov 18, 2016
12:30 p.m.
2 p.m.
Lopata Hall, Room 101

"Performance Analysis of Linux IP Forwarding"

Adam Drescher Advisor: Patrick Crowley

Anyone familiar with the networking subset of the Linux kernel knows that it is a complete maze of configuration options. Additionally, in both academia and industry, there is not a clear source on the IP forwarding capabilities of a default Linux kernel. The stark majority of the analyses focus on filling pipes with large packets. However, the key metric to show packet forwarding capability is the number of minimum size packets forwarded per second, as they have the highest possible amount of per packet overhead. Lastly, much of the networking literature does not account for experimental noise, which should be carefully accounted for in such a volatile subsystem of the kernel. In an attempt to remedy these perceived shortcomings, we present a robust performance analysis of the Linux kernel's IP forwarding capabilities. We evaluate the Linux kernel for both 1 Gbps and 10 Gbps links, and our key metric is the number of minimum size packets forwarded per second. Additionally, we hope to carve away at the vast configuration space provided by Linux, and illuminate a small subset of configuration options that actually matter for IP packet forwarding.

"Automatic cinemagraph for urban scenes"

Hang Yan
Advisor: Yasutaka Furukawa

Cinemagraph is a new type of media which combines the static image with infinite looping video. We have developed a system to automatically create cinemagraph renderings from videos capturing urban scenes. We first apply state-of-the-art 3D reconstruction and image warping algorithms to align all video frames into the reference view. The new video segmentation algorithm based on temporal analysis is then used to select regions to animate and create initial cinemagraph rendering. The initial result is then passed to post processing steps for refinement. We demonstrate the effectiveness of the system on various input videos.