> Robert Raszuk > Sent: Friday, May 3, 2019 3:16 AM > > Radu, > > The MPLS in modern DC is none starter purely from technology pov. > > In modern DCs compute nodes are your tenant PEs all talking to rest of the > fabric L3. So if you want to roll MPLS you would need to do that to the > compute nodes. That means that with exact match you will see in MSDCs > millions of FECs and millions of underlay routes which you can not > summarize. Plus on top of that an overlay say L3VPNs for the tenant/pods > reachability. > Well I guess whenever the summarization is used in the pure IP underlay, in MPLS underlay a seamless MPLS boundary would be used so all the underlay routes/FECs would then be contained only to compute nodes.
> Good luck with operating that scale with MPLS forwarding. Besides while > some vendors of the hosts NICs claim support for MPLS they do that only on > ppt. In real life take very popular NIC vendor and you will find that MPLS > packets do not get round robin queuing to kernel like IPv4 or IPv6 but all line > up to a single buffer. > > Only hacking the firmware of the NIC with some other NIC vendor which also > out of the box was far from decent I was able to spread those flows around > so performance of MPLS streams arriving at the compute was acceptable. > Hmm good to know, I wasn't aware of this. I guess this was specific to a certain setup right? https://www.netronome.com/blog/ovs-offload-models-used-nics-and-smartnics-pr os-and-cons/ adam _______________________________________________ cisco-nsp mailing list [email protected] https://puck.nether.net/mailman/listinfo/cisco-nsp archive at http://puck.nether.net/pipermail/cisco-nsp/
