[Feature Request] Support disable_new_netns with Kube-OVN CNI.

[xujunjie-cover]

Background

Kube-OVN can provide tap or vhostuserclinet device for Kata Containers. Kata Containers can used these devices directly.
Kube-OVN support ovs-dpdk (datapath_type netdev).

Status

Detailed introduction link #1922

current data flow like this:

Proposal

Kata containers use tap devices in host NetNS. then data flow like this:

we can reduce network stack jumps.

control flow for using tap device directly:

the part of kube-ovn is same as provide vhost-user.

Need Kube-OVN to do

• create dummy device to transmit network info.
• create multiple vhost-user or tap device for one kata container.
• support use tap or vhost-user device as frist nic.

Need Kata to change

• add support for vhost-user with server type like this. "server=on"

-chardev socket,id=char-6ca95a82dad2acd6,path=/var/run/openvswitch/vhost_sockets/c3851524-a4bb-48bb-8878-98449e988b7e/hugepages/sock,server=on

• add support: multiple nic queues for kata with multiple cpus. get better performance.
• add support tap device in host netNs like this.

-netdev tap,id=network-0,vhost=on,ifname=tap_69adfedcf5c,downscript=no,script=no

Performance

TODO: we are doing test. More details will be shown here later.
some test using vhostuserclinet provided by ovs-dpdk:
ovs: 4c8g

1. ovs-dpdk + dpdk app runing in kata

Tools: dpdk-testpmd
Version: v21.02
Driver: uio_pci_generic
Hugepages: 512 2M
cpu: 1c already set by cpu_manage. (Intel(R) Xeon(R) Gold 6140 CPU @ 2.30GHz)
Env: two kata containers in same node.

Packet Len(bytes)	pps(M)	Throughput (Gbits/second)
64	6.5	3.1
1400	3.93	41

2. ovs-dpdk + general app

Tools: qperf
Hugepages: 512 2M
cpu: 2c already set by cpu_manage. (Intel(R) Xeon(R) Gold 6140 CPU @ 2.30GHz)

we thought the performance limitations is from qperf use cpu.
two kata containers in same node

Packet Len(bytes)	tcp_lat(us)	Throughput
64	10.7	518 Mb/sec
256	9.7	1.36 Gb/sec
512	11	1.83 Gb/sec
1024	10.1	2.25 Gb/sec
2048	14.9	3.19 Gb/sec
4096	16.2	6.35 Gb/sec
8192	17.5	9.8 Gb/sec

two kata containers in different node, physical nic X710 10GbE SFP+

Packet Len(bytes)	tcp_lat(us)	Throughput
64	17	556 Mb/sec
256	17.1	1.67 Gb/sec
512	18.1	2.41 Gb/sec
1024	19.4	3.09 Gb/sec
2048	39.9	7.55 Gb/sec
4096	40.8	8.49 Gb/sec
8192	43.9	8.99 Gb/sec

Related Link:

kubeovn/kube-ovn#1811

[bergwolf]

cc @kata-containers/architecture-committee

[oilbeater]

@xujunjie-cover do you have the comparison data before and after the optimization