Hi Joel, Regrading this question: Do you have any comments on what appears to be the significant increase in complexity on the device performing PSP?
I think the removal of some bytes from an packet header is a common case without any significant increasing complexity for a long time. For example, a Layer-2 switch receives an ethernet frame from a port with a vlan tag, and forwards/switches it to another port without a vlan tag, then it strips the 2 bytes of vlan tag value and 2 bytes of 0x8100. The 4 bytes are in the middle of the ethernet header and the ip header, which otherwise would not need to be stripped if it's the case from a tag port to another tag port. I found this in my book, "Network system design using network processors" by Douglas E. Comer 2004. On page 38 "4.12 Operation and Data Chaining", introduced modern NICs use a technique known as “operation chaining or command chaining” instead of operating on a single large buffer. Which makes me believe that, the removal of some bytes from some packet header (L2 or L3 header), would not cause the 1500 bytes of packet be moved, but instead a very basic function that various packet processors have solved very long ago as a basic function of "header/packet modification". Hope the above understanding is right. Thanks Jingrong ________________________________________ From: spring [spring-boun...@ietf.org] on behalf of Joel M. Halpern [j...@joelhalpern.com] Sent: Wednesday, December 11, 2019 22:15 To: spring@ietf.org Subject: Re: [spring] Is srv6 PSP a good idea Thank you Jingrong for providing some of the other motivations. Two furhter comments. As far as I know, the only savings on the end box is the processing for noticing the SRH, noticing that SL is 0 and there are no relevant TLVs, and then moving on. If the actual end device is not part of the SR domain, I assume that encapsulation would have been used, so I think it is reasoanble to assume that in the PSP case the end device is SR capable. Do you have any comments on what appears to be the significant increase in complexity on the device performing PSP? The question I am trying to get at is about the tradeoff, which needs one to evaluate both sides. Yours, Joel On 12/10/2019 11:13 PM, Xiejingrong (Jingrong) wrote: > I think it's a good idea. > Nothing new, but benefits that people have already said seems notable to me. > > (1) reduce the load of final destination. This benefit can be notable for the > following sub reasons. > (1.1) final destination tends to have heavy load. It need to handle all the > EHs and do the delivery/demultiplex the packet to the right overlay service. > (1.2) example 1, the final destination may need to handle the DOH after the > RH. > (1.3) example 2, the final destination may need to do the assembly of > fragmented packets. > (1.4) example 3, the final destination may need to do AH/ESP after the > Fragmentation Header. > (1.5) example 4, the final destination may need to deliver the packet to the > right overlay service. > > (2) support the incremental deployment when final destination(s) do not > process/recognize SRH. This benefit can be notable for the following sub > reasons. > (2.1) A core router may (fan-out) connected with a big number of low-end > routers that do not support SRH but support tunnel-end/service-demultiplex > function of SRv6. > > Thanks > Jingrong > > -----Original Message----- > From: spring [mailto:spring-boun...@ietf.org] On Behalf Of Joel M. Halpern > Sent: Wednesday, December 11, 2019 10:55 AM > To: spring@ietf.org > Subject: [spring] Is srv6 PSP a good idea > > For purposes of this thread, even if you think PSP violates RFC 8200, let us > assume that it is legal. > > As I understand it, the PSP situation is: > o the packet arrives at the place (let's not argue about whether SIDs are > locators) identified by the SID in the destination address field o that SID > is the next to last SID in the SID list o that sid is marked as / known to be > PSP o at the intended place in the processing pseudocode, the last (first) > entry in the SRH is copied into the destination IPv6 address field of the > packet > -> The SRH being used is then removed from the packet. > > In order to evaluate whether this is a good idea, we have to have some idea > of the benefit. It may be that I am missing some of the benefit, and I would > appreciate clarification. > As far as I can tell, the benefit of this removal is that in exchange for > this node doing the work of removing the SRH, the final node in the SRH does > not have to process the SRH at all, as it has been removed. > > I have trouble seeing how that work tradeoff can be beneficial. > Removing bytes from the middle of a packet is a complex operation. > Doing so in Silicon (we expect this to be done in the fast path of > significant forwarders as I understand it) requires very special provision. > Even in software, removing bytes from the middle of a packet requires > somewhere between some and a lot of extra work. It is distinctly NOT free. > > In contrast, we have assumed that the work of processing SRH itself is > tractable, since otherwise all of SRv6 would be problematic. So why is this > necessary. > > Yours, > Joel > > PS: Note that both the MPLS case and the encapsulation case are very > different in that the material being removed is at the front of the IP > packet. Pop or prepend are MUCH easier than middle-removal (or > middle-insertion). > > _______________________________________________ > spring mailing list > spring@ietf.org > https://www.ietf.org/mailman/listinfo/spring > _______________________________________________ spring mailing list spring@ietf.org https://www.ietf.org/mailman/listinfo/spring _______________________________________________ spring mailing list spring@ietf.org https://www.ietf.org/mailman/listinfo/spring
