Felix Xiaozhu Lin

Felix Xiaozhu Lin 林 小竹

PhD (Rice ‘14); MS/BS (‘08/’06 Tsinghua). Grew up in a small midwestern town in China. Before UVA, six great years at Purdue.

A new instructional OS

A guided journey starting from bare-metal hardware to a fully functional OS capable of running multiple applications, including Mario and DOOM, on a multicore processor.

Github | Paper

LLM (RWKV) on $30 hardware, under 3 Watt.

Model code | Demo code | Paper

https://github.com/user-attachments/assets/3ff2dea0-e8e2-428c-ae50-8fa0e65eb2cf

Whisper streaming on Macbook Air, with 1 sec of delay. Total SoC power ~7 Watt.

Code | Paper

https://github.com/user-attachments/assets/ad59f529-3567-4236-aabb-cca3ae0eeff8

1st LLM workshop at UVA (Oct 19/20 2024)

Prof. Yangfeng Ji and I organized the 1st LLM workshop at UVA.

Fall 2024: We are recruiting

(Recent projects) algo x sys x hw for:

• On-device LLMs & Speech: on low-cost hardware
• Stable diffusion: fast sampling on consumer CPU/GPU
• Various old-fashioned, hardcore OS stuffs (see pubs below)

My group

I led a group of graduate/undergraduate students who are execited by building cool software stuffs.

See our group page for current students and projects.

Virginia is for lovers!

The UVA campus at sunset, with Blue Ridge mountains in the back. Credit: UVA.

Efforts on refreshing OS education (UVA CS4414/CS6456). Principles: experience-based, with modern contents, and fun. Summary: four projects including an Arm64 baremetal kernel, multicore, trusted execution, and filesystem forensics.

The first offering in Fall’20 as CS6456. Well received! Materials; student feedbacks; schedule

Research (Publications)

I care about systems software. I work with a group of students with enthusiasm in systems software. See XSEL.

Our current work bases on two premises:

1. Old kernels (e.g. Linux) are firmware

• Reuse via Replay: GPU replay [ASPLOS’22, Eurosys’23]; Driverlet [Eurosys’22].
• Transkernel: unmodified Linux drivers on microcontroller-like cores (ATC’19 and HotMobile’17).
• Overwatch: safeguarding unmodified file systems for IoT devices (preprint)
• Power sandbox: power awareness redefined (Eurosys’18).

2. New OSes are defined by scenarios

• STI: turbocharged NLP inference [ASPLOS’23]
• Elf: autonomous AI cameras (Mobisys’20, 3-min video, website, slides)
• VStore: managing large archival videos for analytics (Eurosys’19).
• StreamBox: stream analytics with manycores (ATC’17), with 3D-stacked memory (ASPLOS’19), and with Arm TrustZone (ATC’19).

My PhD work was on OS kernels for smartphones, e.g. the K2 project which won an award in ASPLOS’14. I continue to care about mobile/wearables, e.g. deep learning on them (WWW’19 & MobiCom’18).

Doc, code, talks

• Notes on hacking mobile GPUs
• GPUReplay
• Driverlet
• Jun 2020: “OS frontiers in the AI era” For UVA graduate students. slides video
• Mar 2019: Overview slides

Useful

Need a reference letter from me? | How do I determine final grades for undergrad classes? | How do we work with undergrads in research? | To signature requesters | Name in Chinese

We are one-hour from the gorgeous Shenandoah national park. Unlimited outdoor fun. Photo: https://virginia.org