Hardware Friendly Vhost vDPA: Towards an Efficient and Migratable Device Model

on YouTube ↗31m

" summary

Proposes a simplified vhost‑vDPA migration model that saves device state as TLV device‑state blobs (device‑level, work‑queue, feature sub‑states) and a backend resume API plus PCI/admin‑Q amendments to restore hardware vDPA without shadow virtqueue mediation. Shows vendor feedback (NVIDIA, Intel) and compatibility paths to translate old migration strings. Claims single‑vDPA removes software relay overhead while shadowing can drive CPU up to 40% CPU spike under DPDK; dirty‑tracking and VirtIO PF admin queue are suggested for efficiency.

AI-written summary · about

" chapters (4)

1. 0:00Start
2. 9:59Migration compatibility burden
3. 15:21Auxiliary device state PCI capability
4. 27:33Shadow VirtQueue v.s. Resume-able

" related talks

• 24mvDPA Support in Linux KernelJason Wang · KVM Forum 2020
• 18mvdpa-blk: Unified Hardware and Software Offload for virtio-blkStefano Garzarella · KVM Forum 2021
• 23mHyperscale vDPAJason Wang · KVM Forum 2021
• 36mShared Virtual Addressing for High Performance Arm Infrastructure PlatformsVivek Kumar Gautam · Open Source Summit + Embedded Linux Conference + OSPOCon 2021
• 21mVirtio hardware accelerators and other newsJens Freimann · DevConf.CZ 2019
• 16mUnifying KVM API for protected VM and utilitiesIsaku Yamahata · Linux Plumbers Conference 2023
• 22mAMD-vIOMMU: A Hardware-assisted Virtual IOMMU TechnologyWei Huang, Suravee Suthikulpanit · KVM Forum 2020
• 29mPCI Endpoint Subsystem Open Items DiscussionManivannan Sadhasivam · Linux Plumbers Conference 2023