Hi everyone, I am currently a 5th-year Computer Science Ph.D. student at Georgia Tech. My advisor is Vivek Sarkar. Prior to my Ph.D. journey, I earned my bachelor degree from Washington University in St. Louis in December 2019.

My research area is dynamic HPC/Parallel program analysis. My Ph.D. thesis focused on understanding, detecting and fixing data races in parallel programs. I have developed several dynamic debugging tools for OpenMP and HClib. Additionally, my research achieves theoretical breakthroughs and utilizes Large Language Model in industrial settings.

Email: fjin35 at gatech.edu

Research problems

I am proud that my research pushes the boundary of human’s understanding of data races and program behavior.

Data Race Visualization: Explore an interactive visualization tool that demonstrates data races detected in parallel programs. The source code can be found here. This work is published in the 2024 IWOMP, and the paper can be found here.

Determinism: given the same input, how do we guarantee the programs will return the same output? How do we even define “same” here? In my ECOOP paper, I give the formal definition of determinism for task-parallel programs with promises, and prove that data race freedom guarantees determinism for such program.

Happens-before relation: task A happens before task B if and only if in all possible scheduing of the program, task A has to finish before task B starts.

Concurrent access: if two tasks does not have happens-before relationship between them, we say they can happen in parallel (or concurrently). The exact order will be determined by the runtime. In one schedule, task A may happen before task B. In the other schedule, it may be the other way around.

Data race: if two tasks concurrently access the same memory location, and at least one of them perform a write, we define this as a data race. Data race is notorious because it makes program non-deterministic. Imagine task A writes 5 to variable x, and task B writes 3 to it. When the program ends, we have no idea what the value of x will be: 5 or 3, that’s a question.

Dynamic race detection: To detect data races while running the program, we use compiler to “instrument” each read and write in the program. Instrument means we insert a check function after each read and write. This is done on the IR level. When the program is running, the check function records each read and write, and see if a previous access causes a data race with the current access.
My ECOOP paper designs a new race detection algorithm and evalutes the implementation of it. It is sound and precise: meaning it will not report non-existing race or miss any races.