>> >This is due to the tunnel outer header that adds VXLAN + UDP + IP + ETH = >> >50 bytes to the inner packets. When TCP stream is tested between VMs, TCP >> >buffer is segmented to 1500 bytes, but with outer header added, it will be >> >1550 that slightly exceed PHY NIC MTU, and results in an additional IP >> >fragmentation into big + small packet pairs in Hypervisor's IP stack. >> > >> >In my environment configuring VM MTU to the same as PHY NIC MTU (1500) >> >leads to 40% performance drop. >> > >> >Does this answer your question? >> > >> >Best regards, >> >Han >> >> Thank you for your reply ,Han. I agree with your analysis and know the >> process , My confusion is, whether to have clear documentation or >> rules that PHY or VM NIC MTU need to be modified when we use >> >> ovs vxlan to communication ? >> >> >> if needed, we should increase PYH NIC MTU or decrease VM NIC MTU ? >> > >This problem exists also on other tunnel protocols such as GRE. We should >always increase PHY NIC MTU when possible. > >In fact, MTU specified by VM doesn't make any sense in a virtualized >environment. Maybe you can try this patch if you are interested: > >http://openvswitch.org/pipermail/dev/2014-May/040027.html > >You don't need to care about VM MTU setting with this patch, and the best >thing is that it will be much faster even comparing with properly changed MTU.
Thanks again, It seem that the patch can solve my problem, I will read the patch and test it in my environment. Looking forward to the exchange. _______________________________________________ discuss mailing list discuss@openvswitch.org http://openvswitch.org/mailman/listinfo/discuss
