Bring RoCE (RDMA over Converged Ethernet) Capability to Virtio-net Device
summary
Extends virtio-net to provide full RoCE by implementing the RDMA datapath in the hypervisor back end rather than in guest drivers or SR‑IOV passthrough, adding a virtio-net feature bit and new control/completion/send/recv virtual queues exposed as an IB device; workflow covers PD creation, MR pinning and key generation, mapping completion queues into userspace and posting QPs so data bypasses the guest kernel. Upstream patches and a vhost user back end are on GitHub, targets are kernel patches, performance numbers, and features like RDMA over VPC and live migration.
AI-written summary · about
chapters (16)
- 0:00Intro
- 0:33RDMA Introducation
- 2:11RDMA in Virtualized Environments
- 3:36Enable RDMA on virtio-net device
- 4:26Comparision
- 5:34Architecture Overview
- 6:54New Virtqueue Types
- 8:03Device Initialization
- 9:32Control Path - Create PD (Protection Domain)
- 10:25Control Path - Register MR (Memory Region)
- 11:49Control Path - Create CQ (Completion Queue)
- 12:48Control Path - Create QP (Send Queue & Receive Queue)
- 13:54Control Path - Modify QP (RC/UC Type)
- 14:44Data Path - Post Send
- 16:11Data Path - Fetch Buffer
- 18:08Status & Future Work
related talks
18mRoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs
22mMetaRoCE: Meta’s RDMA Transport
20mMetaRoCE: Meta’s RDMA Transport
16mMetaRoCE: From Spec to NIC to Open Source
12mPreview: Networking
20mRDMA at Cloud Scale: The OCI Experience
28mPersistent Memory plus RDMA, new age remote device
18mFileMR: Rethinking RDMA Networking for Scalable Persistent Memory