Dear Behcet

You need to read the scenarios draft to see the use cases they have in mind.

I think some people have concerns at the cost of compressing a header, but 
there seem to be other use cases.

It is up to the authors to build their argument for short addresses I was 
responding to the case they made.

Stewart


> On 3 Feb 2021, at 15:38, Behcet Sarikaya <[email protected]> wrote:
> 
> Hi Stewart,
> 
> Thanks for your analysis.
> I haven't read the drafts you mentioned but I thought that the address size 
> issue was long resolved with the IPv6:
> 
> basically it matters on the wireless medium and this is solved by the 
> so-called ROHC RObust Header Compression, 
> which is adapted by 5G and it works well.
> Wired medium case which seems to be the main focus according to your mail, 
> was considered moot, the routers would be able handle it and on the medium it 
> does not delay things much.
> 
> Behcet
> 
> On Wed, Feb 3, 2021 at 8:09 AM Stewart Bryant <[email protected] 
> <mailto:[email protected]>> wrote:
>  
> Re drafts:
>  
> https://datatracker.ietf.org/doc/draft-jia-scenarios-flexible-address-structure/
>  
> <https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdatatracker.ietf.org%2Fdoc%2Fdraft-jia-scenarios-flexible-address-structure%2F&data=04%7C01%7Ckiranm%40futurewei.com%7C95b5d102feaf4674ab8408d8c7972448%7C0fee8ff2a3b240189c753a1d5591fedc%7C1%7C0%7C637478799262464227%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=yDi0mFnbU60nFC5PJC%2BAAWVIdSMT%2FY8UO0XIiK3J4iI%3D&reserved=0>
> https://datatracker.ietf.org/doc/draft-jia-flex-ip-address-structure/ 
> <https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdatatracker.ietf.org%2Fdoc%2Fdraft-jia-flex-ip-address-structure%2F&data=04%7C01%7Ckiranm%40futurewei.com%7C95b5d102feaf4674ab8408d8c7972448%7C0fee8ff2a3b240189c753a1d5591fedc%7C1%7C0%7C637478799262464227%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=XB9VFEQiaa0ZMjG5BuF%2FPeQnvFcmGgfY0%2Bye4s7CSoA%3D&reserved=0>
> I have a number of comments these drafts, but in this email I would like to 
> focus on the small address proposal. By this I mean addresses that are 
> shorter than 64bits.
>  
> It seems to me that there are 3 operational reasons for small addresses:
>  
> 1) That you might be worried about the amount of packet taken up by the 
> addresses.
>  
> 2) That you might be worried about the amount of energy required to send a 
> packet.
>  
> 3) That you want to use an address that is somehow native to a legacy 
> application.
> 
> … and there are at least to implementation reasons:
> 
> 4) That you might be worried about the amount of memory in the FIB.
>  
> 5) That you might wish to optimise out the FIB hardware. 
>  
> There are two approaches to the case of addresses that are are relatively 
> short, where for the purposes of this discussion I define “relatively short” 
> as an address between 8 and 64 bits in length.
> 
> One approach is to design and implement a new packet type, be that an 
> original design, or the repurposing of a suitable existing non-IP design. If 
> that is what you have in mind it would greatly assist consideration of your 
> work if you published that design in the IETF, or at least pointed to it as 
> an accessible document. That would allow us to debate the properties of the 
> packet and or your address proposal in the context of the packet design. An 
> alternative approach which needs to be considered is to make the FlexIP 
> address a suffix of an existing and well known address type such as IPv6. In 
> such a case by standardising the corresponding IPv6 prefix you may produce 
> implementation simplifications, or alternatively by making it a prefix well 
> known in the domain you construct quite an effective leakage prevention 
> mechanism.
> 
> Consider the IPv6 suffix case. Validating a well-know prefix before invoking 
> the address lookup machinery is a simple efficient process using either one 
> of more compare operators, or some hw technique such as a special register. 
> Certainly we could build hw to look up a small set of well-known prefixes 
> that burns a lot less energy than used in a full address lookup. So that 
> brings us to looking up the suffix and, by definition the table used to do 
> that is small.
> In other words most of the efficiency of doing a short address lookup can be 
> maintained even if the address is the suffix of a longer address provided 
> that the implementation is optimised for this case.
>  
> I think that argument covers much of use cases 3, 4 and 5. This applies to 
> your Indexes 1..EF, F0, F1 and F2 in your propose address design.
>  
> In your F5 case a longest match engine will work by definition on a variable 
> length address, provided it is short enough that sufficient addresses from 
> the primary address space can be deployed to this. Note that you can throw 
> the address length into the longest match engine if you wish and it will 
> simply consume it and if correctly programmed will return the correct result. 
> Thus any of the address definitions that do not contain discontinuous 
> substructure such as the cases with inbuilt segment routing can be looked up 
> in a common hardware address recognition engine without analysis of the first 
> byte.
>  
> Now I think that it is worth looking at case 1 and 2 above and noting there 
> is some applicability to both the short address cases and the segment routing 
> case with short addresses that you propose.
>  
> Of course it is very difficult to do an accurate analysis these cases because 
> the packet design that FlexIP is going to be used with is not referenced by 
> the drafts.
>  
> Let us assume a tiny packet:
>  
> 14B of MAC header (Lora is 13 to 28, Ethernet is 14)
> 8B of UDP
> 2B of payload
>  
> That is 24B + NW layer (the addresses plus the overhead)
>  
> Now consider IPv6 which is pretty minimalist for a connectionless packet 
> apart from the size of its addresses.
>  
> IPv6 is 40B 
>  
> So total of 64B for the packet which is the benchmark since it is the IETF 
> plan of record for most applications.
>  
> If we reduce the addresses to 1B (the minimum in these drafts) i.e. subtract 
> 32 - 2 = 30 so best case is packet of 34B. 
>  
> That is a useful saving 47% which might be important in some specialist 
> applications where bandwidth or radio energy was important, however  much 
> depends on what the practical size of the payload in, and what options or 
> extensions are in the packet to fulfil the  communications needs. So for 
> example if I were to need an additional 20B of packet option and payload the 
> saving is reduced to 36% reducing to 20% saving if 100B were needed. A 20% 
> saving is not worth the cost and complexity of changing the packet.
>  
> So, to understand the benefit of reducing the address size and presumably 
> using it in an as yet undefined packet format it is necessary for the authors 
> to describe the application for small standalone addresses (as opposed to 
> address suffixes) in a lot more detail than is provided in the scenarios 
> document, in particular the size of the transport layer, and the size of the 
> expected payload. Additionally it is necessary that they describe the details 
> of the network layer packet and its MAC environment.
>  
> Once more detail is know, we will know whether there is a case for small 
> native addresses or whether we should focus our attention on how to map such 
> addresses as suffixes of a larger well known address type such as IPv6.
>  
> - Stewart
> _______________________________________________
> Int-area mailing list
> [email protected] <mailto:[email protected]>
> https://www.ietf.org/mailman/listinfo/int-area 
> <https://www.ietf.org/mailman/listinfo/int-area>

_______________________________________________
Int-area mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/int-area

Reply via email to