Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
Learn about XSched, a novel scheduling framework that enables preemptive scheduling across diverse XPUs (GPUs, NPUs, ASICs, and FPGAs) in this 15-minute conference presentation from OSDI '25. Discover how researchers from Shanghai Jiao Tong University address the critical limitation of current XPUs that lack flexible scheduling capabilities needed for multitasking environments with varying application requirements such as priority and fairness. Explore the unified XQueue abstraction that provides preemptible command queue interfaces for scheduling XPU tasks, and understand the multi-level hardware model that optimizes performance on mature XPUs while maintaining compatibility with emerging hardware. Examine the framework's impressive generalizability demonstrated through adaptation to ten different XPUs across various types, brands, and generations spanning seven software platforms, along with implementation of two hardware-agnostic scheduling policies. Review three comprehensive case studies showcasing XSched's effectiveness in achieving diverse scheduling objectives through efficient and flexible preemption mechanisms in real multitasking workloads.
