Weibin Sun

50 South Central Campus Drive
Room 3190
Salt Lake City, Utah 84112

I have graduated my PhD in summer 2014. Now I am working at Google.

I worked at Flux group. My advisor is Robert Ricci.

My research interest is heterogeneous computing systems. Particularly, the GPU+CPU computing in high performance storage, network, and security systems. My PhD thesis (PDF) includes my Snap and GPUstore work, and also a general survey of the system level GPU computing, which covers both applications and techniques. If you want to know about the system level GPU computing applications and optimization techniques, I would suggest read my defense slides (PDF) first. The slides is enough to help you get the ideas, read the thesis only if you really have time and patience.

I have been working two projects on such GPU+CPU system: GPUstore (previously KGPU) and Snap. I am currently working on building a practical verifiable cloud storage system with GPU-accelerated crypto operations. I also work on Emulab, writing automatic switch configuration tools to set up network connections for Emulab experiments.

I am (sort of) a Linux kernel hacker, absolutely C enthusiast, Python lover, algorithm and UAV hobbyist.

I have a Resume. And also a very simple and stupid dotEmacs.

I also have a Github account where all my research and hobby projects are hosted on.


NVIDIA Graduate Fellowship
It is my great honor to accept the 2011 NVIDIA graduate fellowship to fund my research on "Augmenting Operating Systems with GPUs". The Fellowship Announcement

GPUstore: GPU Computing for Fast Storage
This is derived from KGPU. GPUstore builds high throughput filesystem (eCryptfs in Linux) and storage device drivers (dm-crypt, and md software RAID) with heterogeneous GPU+CPU computing acceleration. Functionality such as encryption and software RAID6 coding and recovery can benefit from such acceleration.

The underlying platform, KGPU, provides a simple and efficient framework for Linux kernel level CUDA GPU programming. We have done a simple survey of kernel level GPU applications in KGPU white paper.

Snap: Fast and Flexible Packet Processing With GPU
Snap is a heterogeneous packet processing platform built on top of Click modular router. We extended Click to Snap by adding wider packet processing pipeline, multithreaded zero-copy packet I/O (based on netmap), batched GPU+CPU packet processing, and multicore NIC RSS support, etc. Snap reaches 40Gbps line rate complex packet processing such as IDS and SDN forwarding, at 128B packets on a very old Core i7 930 + GTX 480 machine with four 10GbE ports, and 30Gbps at 64B smallest packet size.

TagFilesystem: Tag Semantic Integration in UNIX Filesystem Abstraction
Forget about directory, tree structures! All you need is tagging your files and TagFilesystem will manage them for you. Perfect integration into UNIX filesystem model, a path is interpreted as a query in which directory names are tags! TagFilesystem works on Linux.


Old Work on CV and AR

Copyright © 2011-2013 :) Weibin Sun ** Generated by Mou.