GPREEMPT: GPU Preemptive Scheduling Made General and Efficient
summary
Repurposes GPU hardware time slicing and a small driver API to achieve near-instant context-switch preemption, claiming preemption latency within 40 μs while preserving state so non‑idempotent kernels run safely. Implements a timeslice-based yield and a hint-based pre-preemption that overlaps context save with CPU-side data prep. Evaluated on NVIDIA A100 and AMD MI100, showing ~24% average latency increase for latency‑critical tasks when co‑scheduled, and about 88.6% of ideal best‑effort throughput.
AI-written summary · about
chapters (9)
- 0:00GPREEMPT introduction
- 1:01LC and BE GPU workloads
- 2:05GPU kernel limits
- 3:32Preemption trade-offs
- 4:36Leveraging context switches
- 5:41Time slice preemption
- 7:03Optimizing preemption
- 8:49Evaluation setup
- 10:37Summary and future
related talks
36mRuby Supercomputing: Using The GPU For Massive Performance Speedup
42mHow shit works: the CPU
14mFoundational HPC Systems for 2020 and Beyond
21mKeynote: EPI, The European Approach for Exascale Ages. The Road Toward Sovereignty
20mThe Atos Quantum Programme: A QPU in 2023?
39mKeynote: European Processor Initiative & RISC-V
27mComputers for Super-Human AI
01h 06mThe One-Decade Task: Putting std::atomic in CUDA.