Re: [v6ops] new draft: draft-yc-v6ops-solicited-ra-unicast

[Ole Troan <[e-mail]>]

Owen,

>> Even on a flat wired L2 network I don't think it is harmful to default to unicast solicited RAs.
>> 
>> If we take a worst case of 10000 hosts on the L2 network all rebooting/initializing at the same time, we will see:
>> - 10000 DAD probes for link-local address
>> - 10000 MLD joins for the solicited node MC for their link-locals
> 
> As was pointed out earlier, we shouldn’t see MLD joins for LL solicited nodes or any other LinkScoped group, unless I am badly misunderstanding things.

a quick wireshark on the IETF wireless shows lots of MLD for solicited node multicast addresses as well as other link-scoped addresses.

we put that in there under the assumption that we needed to, to deal with MLD snooping switches. I’m not sure that assumption held true.

>> - 10000 RS messages(*)
>> - 1/10000 RA messages
> 
> I think you mean 1 to 10000 because I don’t think fractional messages are possible.
> 
>> - 10000 DAD probes for global addresses (N x 10000 if N prefixes on-link)
>> - N x 10000 mDNS etc type packets
>> 
>> (*) RFC 4861 suggests receiving an RA before an RS is sent, thus under some timing conditions some RS messages might be avoided.
>> 
>> But at best we seem to be talking about saving 20% of the packet during the boot of the 10000 hosts.
> 
> True… However, even if this is an actual concern, we could consider a threshold in PPS where we send an RA Mcast response.
> 
> e.g. if more than 10 RS received in 1 second, send a multicast RA.
> 
> In fact, we could probably get away with delaying an RA response to RS for 250ms. If another RS arrives during that delay, respond multicast, else unicast.
> 
> That probably still eliminates most of the RA traffic, may eliminate many of those 10000 RS messages, and could provide kind of the best of both worlds.
> I’m not sure how hard it would be to implement in silicon, however.

the issue isn’t necessarily link-utilisation.
there are router control planes that would rate-limit 10000 RS messages / second. I agree with Owen and I don’t think we can make a blanket statement that RS replies should always be unicast.

cheers,
Ole