Hi Linda, On 11/13/2012 08:43 PM, Linda Dunbar wrote: > Suresh, > > Comments are inserted below: > >> -----Original Message----- >>> >>> Even though that IPv6 ND uses multicast instead broadcast, IPv6 ND >> does >>> have scalability issues in Data Center when hosts within same subnet >> are >>> spread across different access switches, and when there are lot of >> those >>> subnets. In particular: >>> >>> * when hosts within the same subnet are spread across multiple >>> switches (or GWs in the draft), the ND solicitation still go to all >>> switches which might have the same subnet hosts attached. The traffic >>> across GWs are same as IPv4's ARP. The only difference is that hosts >>> impact is reduced. >> >> This is not true. If the switches are MLD snooping then the packets >> will >> not be sent to irrelevant GWs/switches. >> > > [Linda] For those GWs which has Hosts belonging to a particular subnet, the > multicast for the subnet will reach all those GWs.
The NSs are NOT multicasted to the subnet. They are sent to a solicited node multicast address that is formed from the last 24 bits of the IPv6 address. So only network segments that have nodes with matching 24 last bits in the address would get these messages when MLD snooping switches are present. > > >>> >>> >>> >>> * When host "a" needs to communicate with host "b" in >> different >>> subnet, "a" needs to send "Neighbor Solicitation" to L2/L3 boundary >>> router, (just the same way as IPv4 ARP). >> >> Correct. But only the router gets it. Not all the other hosts in the >> subnet (which is the case for ARP). >> >>> >>> When there are many subnets spreading across many access switches, >> the >>> L2/L3 boundary router has to enable all those subnets on many of its >>> ports/links. >> >> This is purely an implementation issue. I know of several router >> implementations where this is not the case. >> > > [Linda] This is the standard setting. When you have one subnet spanning over > 20 server racks, with each rack reaching the L2/L3 boundary router via > different links, the subnet has to be enabled on all those 20 ports. > > > Why do you say it is implementation issue? Because it is :-). As an example, if you take an fat tree architecture, the number of access switch ports on a subnet is not directly related to the number of ports on the L3 gateway that are on the same subnet. Thanks Suresh _______________________________________________ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
