Introducing UACCE Bus of DPDK
summary
UACCE adds SVA between accelerators and processes and is upstreamed in the Linux kernel (since 5.7) and DPDK 24.03, letting compressors, crypto, DMA and Ethernet devices register as char devices and map doorbell, DS and Q regions into user space. Design focuses on zero-copy DMA, NUMA-aware memory binding and CPU pre-touch; tests show page-fault handling costs ~4x, remote-NUMA hurts ~55% and cross-die ~9.5x, and general workloads see roughly ~50% typical speedup when drivers and kernel support are in place.
AI-written summary · about
chapters (10)
- 0:00Introduction to UACCE
- 1:32What is UACCE framework
- 4:03UACCE design objectives
- 7:04UACCE device interaction
- 11:26UACCE usage scenarios
- 12:33NUMA performance challenges
- 18:00Performance optimizations in DPDK
- 21:10Future directions for UACCE
- 22:35Summary and takeaways
- 23:20Q&A session
related talks
36mBringing vDPA to Life in Kubernetes
24mvDPA Support in Linux Kernel
51mAccelerate containers and VMs networking with vDPA
26mRunning DPDK on Non-Cache Coherent Platforms: Challenges and Solutions
23mHyperscale vDPA
25mDPDK performance optimization for edge computing scenario
31mHardware Friendly Vhost vDPA: Towards an Efficient and Migratable Device Model
21mPreview: Operating Systems