Building GPUVision: A GPU Warp Scheduling Simulator

Warp scheduling determines which groups of threads advance during simulated GPU execution. GPUVision focuses on making those scheduling decisions observable so scheduler behavior can be compared through metrics instead of treated as a black box.

The simulator is written in C++17 and models warp execution, streaming multiprocessor activity, instruction progress, and memory-delay behavior. The goal is a configurable study tool for warp execution and scheduling behavior, not a cycle-perfect model of a physical GPU.

Workload configuration makes it possible to vary instruction mixes and memory delays across benchmark runs. That lets scheduler results be compared under different execution patterns without rewriting the simulation engine.

GPUVision tracks IPC, total cycles, SM utilization, and warp completion time. These metrics give the dashboard enough structure to explain both aggregate scheduler performance and lower-level execution behavior.

More than 14 Google Test unit tests validate core simulator behavior. The tests keep scheduler comparisons grounded by checking the engine logic before performance results are exported.

Benchmark runs export CSV and JSON data so results can be inspected outside the simulator and consumed by the dashboard. Keeping the export formats simple made it easier to debug metrics and build frontend visualizations.

The Next.js dashboard visualizes warp states, SM activity, execution timelines, and scheduler metrics. It turns simulation output into a more readable interface for comparing scheduler behavior across benchmark runs.

The project reinforced how systems work benefits from tight feedback between tests, metrics, and visualization. Clear exports and focused validation made the simulator easier to reason about as features were added.