Hi mcroce, There are spinlock contention problem in the kernel when the packets upcall from a multiple queue netdev with the following patch dpif-netlink: don't allocate per thread netlink sockets. For a vxlan device example the vxlan device recieve the packet from hardware multiple queue. So the packet can be receive on different cores. And these packets did upcalls to the single netlink socket （for vxlan device）. The the spinlock in the wake_up_common will meet the spinlock conention problems and it will leadd all the cpu recieve the packet with higer SI. This can be reporduce with your test but with vxlan device and mutiqueue hardware underlay device. BR wenxu _______________________________________________ dev mailing list [email protected] https://mail.openvswitch.org/mailman/listinfo/ovs-dev
