Hi, All, I am working in container4nfv (old name: openretriever) in opnfv.
Most of work is to integrate k8s, ovs with dpdk, vpp as http://gerrit.opnfv.org/gerrit/openretriever: I developed simple CNI for vhost-user/virtio-user. There are many differences between veth and vhost-user/virtio-user. 1. in veth case, CNI create veth pair for backend and pause container. In vhost-user/virtio-user, CNI need to create vhost-user/virtio-user for POD container instead of pause container. 2. in veth case, POD containers share network namespace with pause container. In vhost-user/virtio-user, ONLY one POD container consume virtio. 3. If we pass unix socket in vhost-user to pause container, it is not easy to retriever that information in POD container. To work around that issue, dummy interface with unix socket ID is created. POD container retriever that information and setup virtio-user with that unix socket id. 4. there is no tcp/ip stack in virtio-user/virtio-user. If following above model, we need to change many CNIs (calico, Contiv, etc) for vhost-user/virtio-user. I just came new idea so that only backend need to be changed and CNIs don't need to change: 1. basic idea is to offload veth (slow path) to vhost-user/virtio-user (fast path) 2. all VHOST message go to veth interface instead of unix socket. Ethertype 0xFFFF is reserved for VHOST message. DPDK need minimal change to support it. 3. if vhost-user/virtio-user is setup, all traffic except VHOST message go to vhost-user/virtio-user. We can also consider the improvement is to aggregate veth interface and vhost-user interface. VPP need some change to support it. 4. application in container can inject traffic from vhost-user to kernel to reuse tcp/ip stack. Is anyone interested in this idea? Thanks, -Ruijing Thanks, -Ruijing _______________________________________________ vpp-dev mailing list vpp-dev@lists.fd.io https://lists.fd.io/mailman/listinfo/vpp-dev
