Hi, Satoru, Thanks for the answers - I think I understand better now. Let me just confirm a few points below...
Satoru Matsushima wrote: > Hi Peter, > > > On Sat, Mar 29, 2014 at 3:32 AM, Peter McCann <peter.mcc...@huawei.com> > wrote: > > >> Ryuji, >> >> After viewing your slides from the presentation you did overnight >> (sorry I couldn't be on the call) I went back and re-read the >> draft-matushima- stateless-uplane-vepc-02 draft. I am still confused >> about a number of things: > > > > Thanks. Let me try to answer your questions. > > > > >> You show in Figure 4, step 15 a Route Update (is this a BGP >> UPDATE?) going from the >> EPC-E to the core network RTR, containing a Destination of UE- prefix >> and a Next-Hop >> of EPC-E address. However, in Section 3.4, you describe the RTR as >> knowing only the >> PDN prefix, which is the same for all EPC-E, and the use of "hot- >> potato" routing to >> deliver the packets to the nearest EPC-E no matter the UE destination >> IP address. >> >> Which one is it? Are the UE prefixes advertised into the core or not? > > > > No, it isn't meant that specific routes to indicate each UEs prefix > are advertised into the core. > I'll try to improve that text in next revision of the draft. Yes please clarify because the current text seems to say that UE prefixes are advertised into the core. > > > >> Assuming for now that the UE prefixes are not advertised into the core, >> but only the PDN prefixes are advertised, then that means that every >> EPC-E must know about every UE session, including the eNB F-TEID for >> every UE in the network, correct? > > > Yes, it's correct. > > >> That's because any one of them at any time could receive a packet >> for the UE from the core. >> This doesn't seem scalable to me. > > > > I agree with you if a EPC-E has whole UE specific routes that exceed > its capacity, it doesn't scale, yes. In the recent presentation > through the webex, Ryuji were trying to explain that it's not intended to do. > Routes contained in EPC-E will be limited/partitioned by operators > policy, such as region, service, population scale, etc., I was a bit confused by the suggestion to partition by region, because there would be no mobility across regions if you partitioned in this way. That's because different regions would use different PDN prefixes. But, I suppose it would be ok to do this if you didn't need to support UE mobility across regions (or if you used OTT mobility such as client MIP for those cases). >> You seem to attempt to address this issue in Section 4.1 when you talk >> about multiple "sets" of EPC-E devices, each one dedicated to a given >> geographic region. > > > Ah, no. Sec 4.1 is intended to explain just scalability issue, and how > to deal that issues with routing techniques in operation. Ok, I guess in the most common case you would have several "slices" of EPC-E, each set serving a different PDN prefix and a different set of UEs. There would be one EPC-E from each slice, each representing a partition of the PDN prefixes, at each EPC-E deployment site between eNBs and core. A given UE's current location would need to be BGP UPDATEd to each of the EPC-E in the slice that covered that UE's PDN prefix. >> It seems to me that each "set" of EPC-E could cover no more than the >> scope covered by a single SGW today, because they each have the same >> amount of state as an SGW. Essentially you have described how to build >> a replicated SGW with failover to different nodes based on the >> re-convergence of BGP after a failure (presumably you could get the >> core network to react to the closure of a BGP TCP session). So I think >> this addresses the problem of fault-tolerance that has been identified >> with the tunnel-based solutions, but not really the scalability >> bottleneck problem. > > The nature of BGP makes easy to do that. I think Sec 3.4 would be > right place to explain that. But I couldn't see that flavor of text in > sec 4.1. Would you point which text in Sec 4.1 makes you confuse? It was the text in the penultimate paragraph that talked about partitioning by region. If you do that, there is no mobility across regions, right? But if you partition by PDN prefix (sets of UEs) then you can have a whole stack of EPC-E at each deployment site, covering the entire population of UEs. >> In fact, if you consider mobility from one "set" to another, if you >> want to keep the >> UE's IP address, you would need to broadcast the same set of PDN >> prefixes from all >> sets of EPC-E. In fact this would mean that all EPC-E throughout the >> network, even >> if they are in different "sets", need to be prepared to handle >> packets for any UE >> and so they ALL would need the eNB F-TEIDs for ALL UEs. Please tell >> me where I have >> made a mistake. > > > > No, an EPC-E just only receives packets from v6 core network toward > UEs that routes installed into EPC-E. Because of that an EPC-E should > advertise aggregated routes only for that includes its downstream UEs. > When the EPC-E advertises whole routes to the core as you explained, > yes I agree with you that won't be scalable. But it would depend on > EPC-E capacity and size of UE population in the network. Ok, so each EPC-E just serves a slice (set of PDN prefixes) of the UE population, right? There is no need to put all UEs on all EPC-Es. -Pete > > cheers, > --satoru _______________________________________________ dmm mailing list dmm@ietf.org https://www.ietf.org/mailman/listinfo/dmm
