XSched: Preemptive Scheduling for Diverse XPUs

XPUs, such as GPUs, NPUs, ASICs, and FPGAs, lack flexible scheduling capabilities, failing to meet rich application requirements (e.g., priority and fairness) in multitasking environments. This paper presents XSched, a scheduling framework that enables preemptive scheduling on diverse XPUs with flexible policies. XSched provides unified interfaces for scheduling XPU tasks through a preemptible command queue abstraction (XQueue). The key challenge in implementing the abstraction is adapting to XPUs with diverse and evolving hardware capabilities and software stacks. XSched proposes a multi-level hardware model that enables mature, advanced XPUs to achieve optimal scheduling performance while maintaining compatibility with emerging, wimpy XPUs. To demonstrate the generalizability of XSched, we adapted it to ten XPUs of different types, brands, and generations across seven software platforms and implemented two hardwareagnostic scheduling policies. We further evaluated XSched through three case studies of multitasking workloads on XPUs. XSched effectively achieves various scheduling objectives using its efficient and flexible preemption mechanisms.