Closed eyakubovich closed 1 year ago
Nothing special. It gets enabled/disabled with
# ethtool -k $NIC tx-udp_tnl-segmentation off
It's been enabled by default on all the NICs we've experimented with, including the Intel ones you linked to.
@jeremyeder That's great news. I now wonder what happens if flannel is running inside a VM. How does VXLAN play with already virtualized NIC w.r.t. to offloads.
If you terminate inside the guest, then no, you won't use the offload. virtio_net doesn't support VXLAN-offload. AFAICT vmxnet doesn't either. If you terminate on the host and send decap'd traffic into a guest, then you're OK.
A team mate of mine wrote a whitepaper detailing what OpenStack Neutron does with VXLAN-offloads: https://access.redhat.com/articles/1749113
@jeremyeder Is this the same as this? https://github.com/coreos/flannel/pull/397
@philips nope.
We need NIC-agnostic vxlan offload...until then, 10G is a pipe-dream in the public cloud.
@jeremyeder I'm using ethtool -K flannel.1 tx off
to disable offloading, somehow it got re-enabled (maybe because of flannel restart? )
What can I do ?
# check
ethtool -k flannel.1 | grep tx-checksumming
# disable it
ethtool -K flannel.1 tx off
This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.
Some NICs support VXLAN offload (http://www.intel.com/content/dam/www/public/us/en/documents/technology-briefs/overlay-networks-using-converged-network-adapters-brief.pdf) but I don't know what needs to be done (if anything) to enable it with flannel.