Re: [pim] pim wg adoption call for draft-chen-pim-srv6-p2mp-path-01

["Acee Lindem (acee)" <[e-mail]>]

Hi Gyan,

From: pim <[e-mail]> on behalf of Gyan Mishra <[e-mail]>
Date: Thursday, January 28, 2021 at 8:48 AM
To: "Jeffrey (Zhaohui) Zhang" <zzhang=[e-mail]>
Cc: Toerless Eckert <[e-mail]>, "[e-mail]" <[e-mail]>
Subject: Re: [pim] pim wg adoption call for draft-chen-pim-srv6-p2mp-path-01

All,

Huaimo has provided some valuable responses to scaling issues.  One point to note as far as scaling is that most all provider run Jumbo frames 9216 MTU and IPv4 length is 16 bits which can support up to 65,535 size packet.

IPv6 header natively supports a 16 bit header length which is the same as IPv4.

RFC 2765 specifies a hbh options Jumbo payload option that carried a 32 bit length field 65,536 and 4,294,967,295 octets in length. Packets with such long payloads are referred to as

   "jumbograms".

All router and switch vendors today support super jumbo which is payload length between 9000 and 65535 capping today at 9216.

In the future as memory and buffer sizes increase in hardware fixed function ASIC FPGA or programmable NPU  for packet IO processing the sizes will eventually jump to 65535.  The next big step will be jumbo grams.  😀

With SRv6 unlike SR-MPLS traditional MPLS based PTA cannot be leveraged such as RAVP-TE or mLDP, so the options for P2MP MDT is very limited and If you don’t want the source node overhead of IR or SR replication SID, you just have BIER as your only option.

Operators love having options and this draft provisss a valuable option that does not require expensive upgraded to support BFR BIER.

But you will require an even more expensive upgrade to support the SRv6 network function to process the MDT subtree in the segment list.

Acee

I support WG adoption as stated and as this draft is very promising for operators as an SRv6 “BIER alternative”.

Thanks

Gyan

On Wed, Jan 27, 2021 at 9:13 PM Jeffrey (Zhaohui) Zhang <zzhang=[e-mail]<mailto:[e-mail]>> wrote:
Hi Huamo,

“For a network with 4k nodes”, how large could a multicast tree be, and with draft-chen, how do you encode that tree in packets? How many copies would you send if each copy is only for a sub-tree?

A large network can be divided into BIER sub-domains, each with a smaller bitstring. If you use tunnel segmentation then you can have efficient replication all the way through (though the segmentation points will have per-tree state). If you don’t want to use segmentation, the ingress can tunnel one copy to each sub-domain and inside each sub-domain you can have efficient replication.

Jeffrey

From: Huaimo Chen <[e-mail]<mailto:[e-mail]>>
Sent: Tuesday, January 26, 2021 10:39 PM
To: Jeffrey (Zhaohui) Zhang <[e-mail]<mailto:[e-mail]>>; Jeffrey (Zhaohui) Zhang <zzhang=[e-mail]<mailto:[e-mail]>>; Bidgoli, Hooman (Nokia - CA/Ottawa) <[e-mail]<mailto:[e-mail]>>; Stig Venaas <[e-mail]<mailto:[e-mail]>>; [e-mail]<mailto:[e-mail]>; 'Toerless Eckert' <[e-mail]<mailto:[e-mail]>>
Subject: Re: [pim] pim wg adoption call for draft-chen-pim-srv6-p2mp-path-01

[External Email. Be cautious of content]

Hi Jeffrey,

    Thanks for your further comments.
    My responses are inline below with prefix [HC4].

Best Regards,
Huaimo
________________________________
From: Jeffrey (Zhaohui) Zhang <[e-mail]<mailto:[e-mail]>>
Sent: Monday, January 25, 2021 10:08 PM
To: Huaimo Chen <[e-mail]<mailto:[e-mail]>>; Jeffrey (Zhaohui) Zhang <zzhang=[e-mail]<mailto:zzhang=[e-mail]>>; Bidgoli, Hooman (Nokia - CA/Ottawa) <[e-mail]<mailto:[e-mail]>>; Stig Venaas <[e-mail]<mailto:[e-mail]>>; [e-mail]<mailto:[e-mail]> <[e-mail]<mailto:[e-mail]>>; 'Toerless Eckert' <[e-mail]<mailto:[e-mail]>>
Subject: RE: [pim] pim wg adoption call for draft-chen-pim-srv6-p2mp-path-01


Hi Huaimo,



Let me pull two points to the top of this email.





[HC3]: This is for the case where the overhead in a BIER-TE packet is 10k bits

when one BitString without any set is used. For the same overhead, how many links

can be encoded.

Zzh3> BIER-TE won’t use 10k bits in the packet. It will use a finite number of bits (e.g. 512 bits or  1024 bits) with multiple sets, just like you break the entire tree into different sub-trees. The comparison is, say with 1024 bits as encoding space, how many nodes/links can each solution encode. I don’t see how draft-chen can do better.

[HC4]: For a network with 4k nodes, there can be about 32k links and 64k BitPositions.
When we use 1024 bits with multiple sets in the packet header,
the number of sets and the size of the set need to represent 64k BitPositions,
but 1024 bits contains a sub-tree that can be encoded in 1024 bits.
If the size of the set is 64 bits, the number of sets needs to be 10 bits.
Each set is encoded using 74 (64 bits for BitString + 10 bits for SI) bits.
1024 bits can contain 1024/74 = 13 sets.
For a sub-tree traversing 13 links with BitPositions in 13 sets,
1024 bits encodes 13 links of this sub-tree in BIER-TE.
With a good SID compression, the size of compressed SID can be 21 bits
(12 bits for node IDs, 3 bits for branches, 6 bits for number of SIDs),
1024 bits can encode 1024/21 = 48 links/nodes of a sub-tree in draft-chen.



[HC3]: It should be a new forwarding scheme for SRv6 replication segment.

Under the network programming, replication SID is a new type of SID.

It will have a new program/procedure for the behavior of this new SID

even though it is “like” existing MPLS P2MP.

Zzh3> The general SRv6 network programing idea is that an IPv6 address is broken into the locator part and func/arg part. The locator part gets traffic to a node and the func/argc part is looked up to decide what to do next. SRv6-P2MP follows that perfectly, so *nothing is new*. When I said it is “like” existing MPLS P2MP, I should also have said that it uses *existing* SRv6 mechanism.

Zzh3> Jeffrey



From: Huaimo Chen <[e-mail]<mailto:[e-mail]>>
Sent: Monday, January 25, 2021 11:45 AM
To: Jeffrey (Zhaohui) Zhang <[e-mail]<mailto:[e-mail]>>; Jeffrey (Zhaohui) Zhang <zzhang=[e-mail]<mailto:zzhang=[e-mail]>>; Bidgoli, Hooman (Nokia - CA/Ottawa) <[e-mail]<mailto:[e-mail]>>; Stig Venaas <[e-mail]<mailto:[e-mail]>>; [e-mail]<mailto:[e-mail]>; 'Toerless Eckert' <[e-mail]<mailto:[e-mail]>>
Subject: Re: [pim] pim wg adoption call for draft-chen-pim-srv6-p2mp-path-01



[External Email. Be cautious of content]



Hi Jeffrey,



    Thanks for your further comments.

    My responses are inline below with prefix [HC3].



Best Regards,

Huaimo



________________________________

From: Jeffrey (Zhaohui) Zhang <[e-mail]<mailto:[e-mail]>>
Sent: Sunday, January 24, 2021 11:26 PM
To: Huaimo Chen <[e-mail]<mailto:[e-mail]>>; Jeffrey (Zhaohui) Zhang <zzhang=[e-mail]<mailto:zzhang=[e-mail]>>; Bidgoli, Hooman (Nokia - CA/Ottawa) <[e-mail]<mailto:[e-mail]>>; Stig Venaas <[e-mail]<mailto:[e-mail]>>; [e-mail]<mailto:[e-mail]> <[e-mail]<mailto:[e-mail]>>; 'Toerless Eckert' <[e-mail]<mailto:[e-mail]>>
Subject: RE: [pim] pim wg adoption call for draft-chen-pim-srv6-p2mp-path-01



Hi Huaimo,



Please see zzh2> below. I trimmed some text.





Zzh> Plus the uSID-Block-ID, for each SRv6 SID that is needed? More below.

[HC2]: The size of compressed SID (u bits) is the average size of

a compressed SID. If compressing method uses uSID-Block-ID, it is considered

in the average size. There are a few of methods for compressing SIDs.

They may have different compression rates.



Zzh2> Before WG adoption, the draft needs to have details spelled out with one of the compression methods. I don’t think it scales even w/ compression, but it’s a non-starter at all w/o compression.

[HC3]: The details about one compression method is posted in one of my

previous responses.

I believe that it is scalable with compression. The details in the example

of one compression method illustrated, on average, one link on a multicast

tree uses 32 bits. This can be improved to 20 bits per link.

It can be a starter without compression in the following way.

For a P2MP/multicast path/tree, the tree can be "split" into multiple sub-trees

such that each of the sub-trees can be encoded in the segment list of the

finite size.



Suppose that the size of SI and BitString are s and b bits respectively, and

the maximum number of possible BitPositions is M bits, the overhead of BIER-TE

is M bits if a maximum BitString is used,

(in this case, M may greater than N*u. For example, when M = 10k, u = 40,

M > N*u for N < 10k/40)

up to N*(s+b) if bit sets are used.

(in this case, N*(s+b) may greater than N*u. For example, when b = 64, s = 10,

u = 40, N*(s+b) > N*u )



zzh> Do you mean “M = 10, 000”?

[HC2]: This is an example value for M. For a network with 10k links, there are 20k bitpositions.

On average, the maximum bitposition is around 10k.



zzh> The overhead consideration should be on a per-packet base. How much of a sub-tree can you encode in *one* packet with a header of a certain size? Your calculation does not seem be for that (and I have trouble following it).

[HC2]: With overhead M bits in a packet, M/u links can be encoded.



Zzh2> I assume you won’t have 10k bits in a packet for encoding the (sub-)tree, right? So the key is how many bits in one packet you can use for encoding the (sub-)tree, and how many leaf/replication nodes you can fit in.

[HC3]: This is for the case where the overhead in a BIER-TE packet is 10k bits

when one BitString without any set is used. For the same overhead, how many links

can be encoded.



Zzh> Additionally, the text seems to assume/imply that there is a continuous block of uSIDs. But with 40-bit uSIDs, you can only put two uSIDs in an SRv6 SID? How large would the segment list be if you want to encode a reasonably sized sub-tree?

[HC2]: "assume" is used to name a variable such as "assume" that

the size of compressed SID is u bits. Here u is a variable for the

the size of compressed SID on average. Some example values for the

variables are given. There seems no imply.

There are a few of methods for compressing SIDs. The method using uSIDs

is one of them. The size of the segment list for a sub-tree is u*N bits,

where N is the number of links on the sub-tree.



Zzh2> How about putting some uSID details in the draft as I mentioned in the other email?

[HC3]: We will add some details in the draft.

3. If this did not involve new forwarding scheme (i.e. new hardware) one could argue that multiple solutions could be developed. But given that this does need new hardware and does not do better than alternatives (e.g. BIER-TE), why bother.

[HC]: It seems that both SR-P2MP and BIER-TE are involved with new forwarding schemes.

Zzh2> SR-P2MP with MPLS does NOT involve new forwarding scheme at all. It’s the same as existing mLDP/RSVP-TE P2MP in the forwarding plane. For SRv6 replication segment, it’s also “like” existing MPLS P2MP in that part of the IPv6 address is used as lookup key just like a label (and that is not different from the SRv6 VPN).

[HC3]: It should be a new forwarding scheme for SRv6 replication segment.

Under the network programming, replication SID is a new type of SID.

It will have a new program/procedure for the behavior of this new SID

even though it is “like” existing MPLS P2MP.

Zzh2> The scheme in draft-chen is so much different - it encodes a sub-tree and forwarding is certainly more complicated. If it does scale well, then  it is worth it even as a new forwarding scheme, but that has not been concluded.

[HC3]: Under the network programming, the behavior of the forwarding

is a new program/procedure, which uses existing encap with the SIDs of a

sub-tree. It seems not that complicated.

Zzh2> BIER-TE is a new forwarding scheme, but it is by far the best per-tunnel TE solution for multicast that does not have per-tunnel state inside the network, and it went through due scrutiny. Draft-chen needs to be scrutinized as well, and it needs to show that it does better to proceed as yet another new forwarding scheme.

[HC3]: Whether it is the best depends on a number of factors, some of them

are explained in previous responses.

Zzh2> Jeffrey

Zzh> Because of new forwarding schemes, BIER went through the process of forming a new WG, and BIER-TE was originally on the experimental track. More importantly, my earlier point is that the new scheme should scale better (at least as well) for it to be worthy.

[HC2]: The new forwarding scheme (or say forwarding behavior) in our draft

can be defined in a small program/procedure in the network programming. This

is very different from that in BIER, where a new WG is formed.

We will update the draft to address your concerns and make it clearer.

Your earlier point is responded in my previous response.



Zzh> Jeffrey



Juniper Business Use Only



Juniper Business Use Only



Juniper Business Use Only



Juniper Business Use Only



Juniper Business Use Only


Juniper Business Use Only


Juniper Business Use Only


Juniper Business Use Only
_______________________________________________
pim mailing list
[e-mail]<mailto:[e-mail]>
https://www.ietf.org/mailman/listinfo/pim
--

[Image removed by sender.]<http://www.verizon.com/>

Gyan Mishra

Network Solutions Architect

M 301 502-1347
13101 Columbia Pike
Silver Spring, MD