Juggler: A Dynamic Warp Scheduler for GPGPUs

General-Purpose Graphics Processing Units (GPGPUs) accelerate various applications, ranging from graphics rendering to scientific simulations and machine learning (ML). The warp scheduler is a critical element in GPGPUs. It needs to manage thread execution in a way that optimizes performance as well as efficiency. However, different applications, and even different phases of the same application, have varying performance requirements. Hence, fixed static scheduling algorithms are suboptimal. This brief introduces a dynamic warp scheduler named Juggler, which dynamically switches between different scheduling strategies at runtime based on phase-specific performance characteristics. Juggler predicts and switches to the most suitable scheduling mode at runtime while improving the overall performance. Evaluations demonstrate that Juggler improves performance by up to 24.84%, with an average gain of 4.66% over state-of-the-art schedulers. These results emphasize the capability of dynamic scheduling to unlock even greater computational power of the GPGPUs.