Machine learning model optimization and inference infrastructure

Implemented a novel data plane network verification algorithm in C++ from scratch, based on a recent academic paper. Network verification models network topology and packet forwarding, and checks for the existing of unwanted behavior, such as loops and blackholes. Drastically increased the performance of network topology modelling and invariant checks (loop) by 500+ times, compared with existing Python tools, enabling real-time verification of TikTok’s expansive global network. Also Implemented incremental rule checking, allowing network configuration to be updated on the fly.

I worked on a software product for FPGA placement, a phase of FPGA physical synthesis. I implemented GPU acceleration on it using CUDA and shortened its runtime by 7 times on average. Apart from improving its performance, I simplified its development, porting the algorithm to a PyTorch-based framework, framing this nonlinear nonconvex optimization problem as training a neural network. I used C++ & CUDA extension on Python to speed up critical segments while maintaining low code complexity. I also extended the algorithm to consider clock network routing resource constraints with a quadratic penalty, considering global placement convergence and design legality. Furthermore, I expanded the algorithm to include new cell types, allowing it to process a non-academic real-life industry architecture and substantially improved its performance. No academic FPGA placement software ever contained these features, so my work resulted in a publication in a top IEEE journal.