Modern machine learning workloads rely on both compilers and runtime systems to schedule work efficiently on GPUs. Static compilation approaches can produce highly optimized schedules when workloads are fixed, achieving near-ideal utilization. However, these schedules are often brittle: small variations in workload characteristics, such as timing differences between stages or the dimensions of the matrices being multiplied, can introduce inefficiencies such as idle time or pipeline bubbles. In contrast, dynamic schedulers adapt to runtime conditions but typically operate with limited global structure, leading to suboptimal overall performance. This project will explore a middle ground between these approaches by investigating whether it is possible to precompute a small set of structured scheduling strategies and adapt among them at runtime.

I am interested in understanding how systems and compilers interact to achieve high performance on modern hardware. Through this project, I hope to deepen my understanding of scheduling, GPU execution, and compiler design, while gaining experience building and evaluating systems at this interface. I also aim to further explore my interest in performance engineering and compiler research.