Nixie: Efficient, Transparent Temporal Multiplexing for Consumer GPUs
Talkat 20th USENIX Symposium on Operating Systems Design and Implementation (OSDI '26) · 2026-07-13..07-15·
summary
Nixie is a system service that temporally multiplexes consumer GPUs by inserting a transparent CUDA shim per application and a centralized daemon that co-manages memory placement and kernel launches. It enforces four-tier storage (GPU, pinned host, pageable host, disk) with explicit budgets, pre-planned bulk transfers using 2 MB blocks, and an MLFQ-inspired scheduler that detects idleness from sync points. On a 132 GB RTX 5090 Nixie hits 18–20 GB/s bi‑directional PCIe, reduces CPU pinned memory by up to 66.8%, and speeds interactive code completion up to 3.8x versus UVM baselines.
AI-written summary · about
chapters (15)
- 0:00Nixie and local ML workloads
- 1:02Transparent GPU sharing challenges
- 2:05Memory constraints on switching
- 3:53Efficient GPU sharing requirements
- 5:18Nixie system architecture
- 6:20Transparent switching mechanism
- 7:24Multi-tier memory management
- 8:47Proactive memory transfers
- 9:29Hint-free scheduling
- 10:52Evaluation methodology
- 11:33Switching speed results
- 11:54Memory efficiency and pinning
- 12:39Multi-workflow comparison
- 13:23Latency and throughput
- 14:05Nixie benefits and open source
related talks
38mExploring Distributed Caching for Faster GPU Training with NVMe, GDS, and RDMA
14mServing Frontier MoE Models at the Lowest Token Cost
36mWhy Are GPUs (Not) Fast - A Trip Through the Driver Stack
23m1.2 Tb/S Packet Processing of Radio Telescope Data on a GPU
24mImproving GPU Utilization and Accelerating Model Training with Scheduling Framework and NRI
20mMemory Harvesting in Multi-GPU Systems with Hierarchical Unified Virtual Memory
43mRealtime performance at scale (...)
30mBeyond the DC Walls: Building a Nationwide GPU Fabric With KubeVirt and Wide-Area…