Projects

A curated list of the projects (excluding research projects) I’ve done.

Cesame: A C-like programming language with First-class Functions and More

Cesame is a C-like programming language that features native support for first-class functions. Cesame also supports non-primitive types such as String, Array, Struct, Socket, and File. Cesame abstracts away pointers and stores most data on the heap, except for primitives and references to objects, similar to Java.
The compiler is written in OCaml with C libraries and targets LLVM IR (phases of Cesame compiler illustrated on the left).

FPGA-based Network Switch: A customizable FPGA-based network switch that supports 4 ingress/egress ports.

daFPGASwitch is a customizable FPGA-based network switch that comes with the following components: A MAC-to-port lookup table unit, per-ingress-port virtual output queues (VOQs), a queue management unit that manages the mapping between packets stored in memory and the VOQ, a scheduling unit that controls the dequeuing decision (which packets get sent to the switching fabric), and a crossbar switching fabric. Different patterns of packet traffic are generated with software to test the performance of different scheduling algorithms.

(MyDB) x (MyMake with cache) in C++

Source Code available upon request. A simplistic database that stores structs in binary form, which can be written to/read from with Writer/Reader, along with a customized mymake program that mimics and extends GNU make’s functionality by adding a cache layer. Parsing MyMakefile and storing target-rule pairs in .mymake.cache file (with MyDB mentioned above) to avoid RegEx processing, variable mapping, and implicit rules generating overheads.

Selected Projects on Parallel Computing

Accelerating Poisson Equation Solving with CUDA

GitHub repo here. In this project, I aimed to optimize solving the 3D Poisson equation using CUDA by parallelizing computations on the GPU. Key achievements include:

• Optimized binary tree parallel reduction up to the warp level using techniques like sequential addressing and warp unrolling.
• Implemented Jacobi, Red-Black Gauss-Seidel, and Block-wise Gauss-Seidel methods in CUDA.
• Achieved a 26x speedup by tuning hyper-parameters and optimizing memory access with 3D texture memory.

AllReduce with MPI

GitHub repo here. In this project, I implemented the SUM AllReduce operation in MPI, commonly used in deep learning to compute gradient means, using four algorithms: Brute Force, Butterfly, Tree, and Ring AllReduce. These algorithms were optimized to handle bandwidth limitations in complex cross-node network environments.

Time Series Research

During my undergraduate studies, I developed an interest in time series and conducted research on using time series to capture the dynamic landscape of China’s burgeoning financial market. Here are the topics I explored during my research.

Price Discovery & Synergistic Effect: A Probe into the Commodity Futures Market

An Observation of the Chinese Soybean-related Commodity Futures Market from a Time Series Perspective. We found these two key phenomena in the commodity futures market with impulse response analysis, SVECM model, and causality test.

• Price discovery: Co-integration of spot and futures prices.
• Synergistic Effect: the synergistic relationship between related agricultural products.

Monetary Policy Shocks with SVAR and DSGE: A Comparison between China and the US

Our analysis using SVAR and DSGE models found that higher interest rates in China create positive short-term and negative long-term effects on GDP, while curbing inflation. In contrast, traditional monetary policy struggles with stagflation in highly developed economies like the US.

RegexHub

Github repo here. A platform to find, share and test regular expressions (with convention in Ruby) in one place. Written in Ruby, powered by Ruby on Rails and deployed to Heroku.