Re: [Detnet-dp-dt] detnet LSPs and PWs

[Stewart Bryant <[e-mail]>]

On 15/02/2017 05:45, Norman Finn wrote:
> There is certainly a problem processing sequence numbers at very high 
> speed.  Unless/until the hardware is able to do that, this technique 
> probably won't be used at really high speed.  You can't do what you 
> can't do. In the meantime, it can be very useful for lower speeds.

If you can force the traffic to arrive at the same physical interface, 
for example by SR or RSVP-TE you are better off. I am not sure how this 
scales because you have to have the s/n in RAM that the forwarder can 
get to quickly, but this would be OK at the egress T-PE since you are 
bounded on the number of PWs, although I suppose that could be # ports * 
# VLANs

Depending on the PE architecture we might be able to do egress 
processing, although that is not a common feature in the platforms I 
have looked at in the past.

>
> Discarding out-of-order packets doesn't makes the whole exercise 
> pointless, but does decrease its utility.  There are two major classes 
> of use cases known:
>
>  1. "Intermittent flows": Where the difference in latency between 
> paths is smaller than the interval between transmissions.  This is a 
> typical case in machine control applications.  In this case, the 
> long-path repeated packet is received before the next short-path 
> packet.  Discarding out-of-order only has the usual issues with reset 
> transmitters.
>
>  2. "Bulk flows":  Where there are several packets more in flight 
> along the slow path than the fast path.  This is a typical case in 
> super high-definition video applications. Tossing out-of-order packets 
> makes the packet replication/elimination pointless; The object is to 
> receive every packet, even if out-of-order.

So in case 2, it is possibly better  to deliver up all the packets and 
let the device do the discard. In video there is usually a sequence 
number system that would permit that, assuming it was nit confused by 
the abnormally high number of duplicates.

Stewart


>
> -- Norm
>
> ------------------------------------------------------------------------
> *From:* Detnet-dp-dt [[e-mail]] on behalf of 
> Stewart Bryant [[e-mail]]
> *Sent:* Monday, February 13, 2017 4:56 AM
> *To:* Andrew G. Malis; Loa Andersson
> *Cc:* [e-mail]
> *Subject:* Re: [Detnet-dp-dt] detnet LSPs and PWs
>
> Talking of corner residing elephants how to make sequence number 
> processing atomic at the egress detnet-T-PE?
>
> This is not a problem at the S-PEs because it does not matter if two 
> copies get through but it is critical at the egress T-PE.
>
> Also what do you plan to do if a later packet overtakes an earlier 
> one? Presumably declare all the late delivery packets as "lost" rather 
> than attempt to re-order.
>
> - Stewart
>
>
> On 13/02/2017 10:55, Stewart Bryant wrote:
>>
>> Hi,
>>
>> I was given some background information on your thoughts on 
>> detnet-PW, and pass on my thoughts in response.
>>
>> I think NSP issue  is a red herring. NSP can be NULL.
>>
>> An S-PE does no NSP, although in any case I suspect that you may need 
>> some processing function at the detnet-S-PE - see below.
>>
>> The underlying DETNET PW is an SS-PW in the diagram I was shown, in 
>> that the PW label is the same end to end, although of course it need 
>> not be, you could have an equivalence label set and run pure MS-PW. 
>> Indeed when you have multiple administrations you would like them to 
>> be different for administrative purposes (that is why we designed 
>> MS-PW like that).
>>
>> So if you create an equivalence relationship in the egress-PE, i.e. 
>> two entries in the global label table pointing to the same duplicate 
>> suppressor (sequencer), then you could use regular MS-PW for this.
>>
>> If the S-PEs are in the same administrative domain in both ingress 
>> and egress, you can also use a single label value on egress and on 
>> egress since you can give them the same label mappings, i.e. they 
>> have identical swap instructions programmed into the L-FIB.
>>
>> We don't have NSP at the S-PE's in the current PW architecture, in 
>> the data-plane it is essentially a simple MPLS LSR, swapping PW 
>> labels and forwarding the packet on a new LSP. What you will almost 
>> certainly want to do is to have the ability to replicate at nodes at 
>> the S-PEs, and that is new functionality.
>>
>> An approach I would look at is as follows:
>>
>> Create a new detnet-T-PE. On ingress this adds the sequence number, 
>> replicates and adds the PW label, which as I said above MAY be next 
>> hop detnet-S-PE dependent. Then it delivers the copies to the 
>> detnet-S-PEs over the LSPs. Now if you have an ECMP path between the 
>> detnet-T-PE and a detnet-S-PE, or you have SR or RSVP-TE available 
>> you can also deliver multiple copies to the detnet-S-PE and take 
>> advantage of the variability of transit time in the MPLS underlay.
>>
>> Now you create a new detnet-S-PE that operates as follows. On it's 
>> ingress side it looks for the first packet at a given sequence number 
>> on this PW (or PW set) and suppresses all future packets on that 
>> sequence number on that PW or PW-set. It then replicates the packet 
>> if required, swaps the PW label (note that it may also use multiple 
>> outgoing labels) and send the packet over the egress LSP set.
>>
>> At the egress T-PE it looks at the sequence number on this PW (or 
>> PW-set), trims all duplicates, applies any required egress processing 
>> and send the packet on it's way.
>>
>> In summary on ingress a detnet-T-PE replicates to multiple S-PEs 
>> using the PW label the detnet-S-PE expects and potentially sends the 
>> packet over multiple paths to the detnet-S-PEs. At egress a 
>> detnet-T-PE looks at the sequence numbers across the detnet-PW set 
>> and selects the first of the sequence number suppressing all others, 
>> and sends the underlying packet on its way. A detnet-S-PE is a back 
>> to back detnet-egress-T-PE and a detnet-ingress-T-PE with a PW label 
>> swap in the middle and no other PW processing.
>>
>> Now for the elephant in the corner of all of the schemes I have seen. 
>> If you have multiple paths to an X-PE, packets will likely arrive on 
>> different line cards. Sequence number co-ordination amongst different 
>> line cards, and at high speed even amongst different ports on the 
>> same line card is a hard problem. Indeed depending on the pipeline 
>> design on the line card, ANY sequence number processing can be hard. 
>> You could mitigate this (at the cost of availability) by requiring a 
>> common ingress port at any detnet X-PE. This would normally require 
>> an RSVP-TE or SR underlay.
>>
>> - Stewart
>>
>>
>>
>> On 13/02/2017 04:11, Andrew G. Malis wrote:
>>> Loa et al,
>>>
>>> To be clear, there’s currently no definition of PWs encapsulated in 
>>> PWs, and while it might be conceptually possible, such as an 
>>> Ethernet PW carried within a SONET/SDH PW, I couldn’t imagine a use 
>>> case for doing it as it’s very inefficient, and I asked Loa if he 
>>> had one. And if you were to do so, each PW in the hierarchy would 
>>> need NSP functionality and real or emulated CE access circuits at 
>>> the endpoints. Also, thinking about it some more, you couldn’t have 
>>> both PWs in the same label stack, since a PW emulates a physical 
>>> circuit. So there would need to be a separate label stack (and MPLS 
>>> LSP) inside the emulated circuit for the outer PW. By definition, PW 
>>> labels terminate a label stack.
>>>
>>> As I haven’t been following Detnet at all, I don’t have the context 
>>> for what you’re trying to accomplish. That said, I’ll take a look at 
>>> the slides and let you know if I have any comments.
>>>
>>> Cheers,
>>> Andy
>>>
>>>
>>> On Sun, Feb 12, 2017 at 8:31 PM, Loa Andersson <[e-mail] 
>>> <mailto:[e-mail]>> wrote:
>>>
>>>     Folks,
>>>
>>>     The mail from Yuanlong mmade me go back and check the PW
>>>     architecture and consult with Andy Malis and Stewart Bryant. So
>>>     I have one more thing
>>>     that we should discuss on "Tuesday".
>>>
>>>     What Yaunlong said was: "IMHO, multiple layers of PW is a break from
>>>     the PWE3 architecture, and all DP/CP/MP things will become more
>>>     complicated."
>>>
>>>     It is correct that multi-layer pw's is problematic, though Andy said
>>>     that "if you have a good use case, you can do it".
>>>
>>>     The problem is that there is a native service processing (NSP)
>>>     at the end of the PW. Multi-layer PWs will only do NSP at one
>>>     level. I think
>>>     we should replace the MS-PW with an LSP. I've added one slide
>>>     (9) and
>>>     change slide 8,9 and 11 in the earlier package. The other slides
>>>     arere
>>>     for reference
>>>
>>>     I want Andy and Stewart to have a chance to review this prior to
>>>     that
>>>     we commit too hard to it. Copied them.
>>>
>>>
>>>     /Loa
>>>     -- 
>>>
>>>
>>>     Loa Andersson                        email:
>>>     [e-mail] <mailto:[e-mail]>
>>>     Senior MPLS Expert [e-mail] <mailto:[e-mail]>
>>>     Huawei Technologies (consultant)     phone: [phone]
>>>     <tel:%2B46%20739%2081%2021%2064>
>>>
>>>
>>
>