HI Linda: I have no problem with the concept that a subnet may span multiple server racks, I expect that to be a common deployment scenario.
In fact the more fragmented the space is the more likely that the cardinality of VM to ARP proxy will approach 1:1 (in which case SARP will offer no state reduction) and the more likely it is that you will need a lot more than 4K VLANs in which case a flat S-tagged bridging domain will not cut it and you would need PBB or some similar technology that enlarged the VLAN space, and while it was at it, summarized C-MACs, which largely solves the state inflation problem SARP purports to solve. So if you fragment to the level you suggest such that SARP supposedly is needed, SARP offers little benefit, and if you do not, you would not need it given current technology as not EVERY MAC in the system would appear in EVERY switch. So I do not see the merit in something that takes away existing L2 capability, and offers a questionable scaling benefit. That's the view from here Dave -----Original Message----- From: Linda Dunbar [mailto:linda.dun...@huawei.com] Sent: Thursday, October 03, 2013 3:42 PM To: Ted Lemon Cc: George, Wes; David Allan I; int-area@ietf.org Subject: RE: [Int-area] Call for adoption of draft-nachum-sarp-06.txt Ted, > -----Original Message----- > > What matters is that an observation has been made that a L2 VPN could > have severe negative behavior characteristics in plausible > circumstances. This is a really strong reason not to use an L2 VPN. [Linda] I understand that many people are still not convinced of the need for hosts within one subnet to be placed in different server racks. For a typical Data Center, there may be many instances of WebServer in one subnet (typical one VLAN), many instances of AppServer in another subnet, and many instances of DBServer in yet another subnet. Being in one subnet doesn't mean that those hosts (or applications) have higher traffic among themselves. One of the main reasons for partitioning hosts to multiple Subnets is for policy enforcement and load balance purposes. The amount of traffic between "WebServer Instances" <-> "AppServer instances" <-> "DB server instances" is magnitude higher than the heartbeat traffic among all instances of one function. Placing hosts based on their Subnets actually creates more across rack traffic than placing instances of different functions (or subnets) based on their traffic patterns. For some DC operators who want to optimize their traffic across racks, placing instances from different subnets in one rack can eliminate a lot of traffic traversing among racks. Linda _______________________________________________ Int-area mailing list Int-area@ietf.org https://www.ietf.org/mailman/listinfo/int-area
