Hi, I am using the TREX DPDK (DPDK version 20.02) traffic generator in a k8s test environment. I use both TREX pods with kernel networking (Linux interfaces using DPDK PF_PACKET PMD driver) and pods with SRIOV/DPDK networking.
One use case is to monitor traffic over long time sending only 1k PPS streams. I have noticed that over time there is some small amount of out of order packets using the Linux interface. It is never the case for the SRIOV/DPDK flows. I also use vpp (FDIO) pod with host-interfaces (i beleive this is also using DPDK PF_PACKET PMD driver) in the setup, also this pod seems to contribute to small amount of reordering. If I use a normal ubuntu pod instead pure kernel forwarding of the packets between two interfaces I do not see any reordering. I think both trex (using "linux interfaces") and vpp (using host-interfaces (DPDK PF_PACKET PMD)) make use of linux packet socket interface. The ununtu kernel networking pod just use normal kernel forwarding. Do you think that maybe the small reordering could be caused by the use of linux packet socket interface? I just try to understand if I have issue in my network fabric or in the DPDK PF_PACKET PMD tools in this case. BR//Håkan
