Currently when dpdk-app is running with the following configuration
l2fwd -n 4 -l 7-10 --master-lcore 7 --vdev=virtio_user0,path=/var/run/openvswitch/bccdd931d211-net1 --vdev=virtio_user1,path=/var/run/openvswitch/bccdd931d211-net2 --no-pci -- -p 0x3 -T 120 --no-mac-updating
When, I check the dpdk-info, I can see this: bccdd931d211-net1 3/4: (dpdkvhostuser: configured_rx_queues=1, configured_tx_queues=1, mtu=1500, requested_rx_queues=1, requested_tx_queues=1) bccdd931d211-net2 4/5: (dpdkvhostuser: configured_rx_queues=1, configured_tx_queues=1, mtu=1500, requested_rx_queues=1, requested_tx_queues=1)
Is it possible to increase the number of queues? If so, should we do it on the ovs in host? Is there a userspace-cni configuration for that?
Currently, when we run the pod, we do not have control on which core they should run.
With PF as ixgbe and VF as VFIO_PCI the setup works. But the issue is with the traffic. All packets are dropped by the PF (kernel driver).
Other scenarios that did not work:
EAL: VFIO support initialized EAL: PCI device 0000:06:10.0 on NUMA socket 0 EAL: probe driver: 8086:10ed net_ixgbe_vf EAL: PCI device 0000:06:10.1 on NUMA socket 0 EAL: probe driver: 8086:10ed net_ixgbe_vf MAC updating disabled EAL: Error - exiting with code: 1 Cause: No Ethernet ports - bye
EAL: PCI device 0000:06:10.0 on NUMA socket 0 EAL: probe driver: 8086:10ed net_ixgbe_vf EAL: Cannot open /sys/class/uio/uio0/device/config: Read-only file system EAL: Requested device 0000:06:10.0 cannot be used EAL: PCI device 0000:06:10.1 on NUMA socket 0 EAL: probe driver: 8086:10ed net_ixgbe_vf EAL: Cannot open /sys/class/uio/uio1/device/config: Read-only file system EAL: Requested device 0000:06:10.1 cannot be used MAC updating disabled EAL: Error - exiting with code: 1 Cause: No Ethernet ports - bye
container fails to start, see this error in /var/log/messages.. LoadConf(): failed to get VF information: "lstat /sys/bus/pci/devices/0000:06:10.0/physfn/net: no such file or directory"
VPP userspace CNI plugin does not support vhostuser interface. VPP userspace CNI plugin support memif, but the POD (dpdk-centos) cannot setup memif, it need vhostuser We have 2 options:
- Explore different CNI for VPP that support vhostuser.
- Consider different pod (container) that can handle memif and still run dpdk-l2fwd application.