Re: [Int-area] Call for WG adoption of draft-templin-intarea-parcels-10

[Tom Herbert <[e-mail]>]

On Tue, Jul 12, 2022 at 7:03 AM Templin (US), Fred L <
[e-mail]> wrote:

> Joel, you are misunderstanding what nodes would be involved in reassembly;
> this would
>
> not be at every single IP layer router in the path. It would only be at
> possibly 0, 1 or 2
>
> adaptation layer middleboxes in the path from source to destination.
>

Fred,

The problem isn't getting this to work on every single router, it's the
requirements imposed on a router even for a single use case. Reassembly is
non-work conserving so that sets a memory requirement on the routers to
buffer packets in reassembly. How much additional memory are routers going
to need for this? Also because it's work conserving, there needs to be
policies and mechanisms on how long to wait in reassembly? And as Joel
alluded to, reassembly is introducing delay in the network, so how does
that interact with congestion control algorithms? For instance, if a node
holds a packet of flow for longer than RTT, then we'll see the source
retransmitting packets not because they're lost but because the network
decided to arbitrarily delay them. So there are a myriad of considerations
for implementing reassembly in the network, and even with the constraints
you've suggested, it is not trivial.

And, then most
>
> likely only at a near-end middlebox very near the destination that happens
> to know the
>
> destination would prefer to receive larger parcels.
>

Hosts already have all capabilities of IP reassembly and TCP reassembly for
many years. This is what host stacks do, and they can do it well. Moving
this functionality into the network doesn't solve any problems the host is
currently facing, but would create new ones.

>
>
> About segment size, I have proof that using segment sizes significantly
> larger than the
>
> path MTU can often produce dramatic performance increases even when
> fragmentation
>
> is intentionally invoked. I also have proof that packaging multiple
> segments in the same
>
> system call can drive performance even higher an without reducing the
> segment size.
>

Yes, but we already get those effects using GRO/GSO and now BiGTCP allows
>64K operations. I don't see how IP parcels can improve upon that.


> IP parcels takes it the logical next step of allowing multiple segments to
> travel together
>
> in the same packet, which may or may not be subject to fragmentation and
> reassembly.
>
> But, let’s not get so hung up on the middlebox question that we forget the
> benefits
>
> for end-to-end.
>

The one material benefit I see is where the source network has a larger MTU
than the destination network. So packets can start big and be fragmented
when they enter the smaller network. This has an advantage over IP
fragmentation in that if a fragment is lost the other fragments are still
useful to the transport layer. As mentioned above, reassembly in the
network isn't feasible, but if some tunneling is used one might be able to
match the functionality so that packets going from a smaller MTU network
into a larger one might be coalesced to take advantage of the larger MTU.
Maybe there's some use cases in satellite communications for this?

Tom


>
> Fred
>
>
>
> *From:* Joel Halpern [mailto:[e-mail]]
> *Sent:* Monday, July 11, 2022 4:02 PM
> *To:* Templin (US), Fred L <[e-mail]>
> *Cc:* [e-mail]
> *Subject:* Re: [Int-area] Re: Call for WG adoption of
> draft-templin-intarea-parcels-10
>
>
>
> No, intermediate reassembly is not an optimization.
>
> First, it is a bad idea.  It is very painful for routers to perform
> reassembly.  They have to burn expensive resources managing such
> attttempted reassesmbly.  It has major cost even if the router decides to
> give up and forward the pieces.
>
> And second, unless one makes some unstated assumptions in the absence of
> such reassembly the sending host will be throttled to the receiving host
> rate.  So the benefit of the entire system is markedly reduced.
>
> Net: we should not adopt this draft.
>
> Yours,
>
> Joel
>
> On 7/11/2022 6:41 PM, Templin (US), Fred L wrote:
>
> Tom,
>
>
>
> > Why would someone put six segments in a parcel if they already have a
> 9K link MTU?
>
> > Why not just send one segment in 9K?
>
>
>
> This is the mindset that we need to overcome. We have had it drilled into
> our heads
>
> that MSS must be the same as the path MTU, but it does not need to be that
> way.
>
> If the MSS is smaller than the path MTU, but we can send multiple segments
> in a
>
> single parcel that more closely approaches the size of the path MTU then
>
> amortization savings are possible.
>
>
>
> >The algorithm isn't the problem, it's supporting new protocols and
> multiple
>
> >checksums in a packet in hardware.
>
>
>
> But Tom, how hard can this be? Instead of running the Internet checksum 1
> time
>
> over N octets of data simply run it M times over N/M octet chunks of the
> data in
>
> succession but still in a single pass. You spoke before of NICs adapting
> to support
>
> TCP jumbograms – if they can do that, why not a very straightforward
> application
>
> of Internet checksum? I haven’t looked at this in a long while, but isn’t
> this also
>
> similar to what UDP-lite did?
>
>
>
> > Either you're trivializing reassembly or maybe you're thinking of some
> new method that
>
> > somehow avoids all the pitfalls and problems we've had with reassembly
> over the years!
>
>
>
> Intermediate node parcel reassembly is really just an optimization to try
> to pass the
>
> largest possible parcels on to the next hop instead of passing many
> smaller ones. It is
>
> really just a concatenation of segments of sub-parcels belonging to the
> same original
>
> parcel. Reordering is unimportant – it is OK to concatenate sub-parcels
> 3,8,5,2 in that
>
> order and without even waiting for any other sub-parcels to show up. The
> application
>
> will simply perceive it as a case of network reordering and the upper
> layer protocol
>
> will do the correct thing with the sequence numbers. AFAICT, the only hard
> requirement
>
> is that the final sub-parcel must not be concatenated as an intermediate
> sub-parcel.
>
>
>
> This stuff will all work, and it will work for the betterment of the
> Internet.
>
>
>
> Fred
>
>
>
> *From:* Tom Herbert [mailto:[e-mail] <[e-mail]>]
> *Sent:* Monday, July 11, 2022 2:57 PM
> *To:* Templin (US), Fred L <[e-mail]>
> <[e-mail]>
> *Cc:* Richard Li <[e-mail]> <[e-mail]>;
> Juan Carlos Zuniga (juzuniga) <juzuniga=[e-mail]>
> <juzuniga=[e-mail]>; [e-mail]
> *Subject:* Re: [EXTERNAL] Re: [Int-area] Call for WG adoption of
> draft-templin-intarea-parcels-10
>
>
>
> EXT email: be mindful of links/attachments.
>
>
>
>
>
>
>
>
>
> On Mon, Jul 11, 2022 at 2:20 PM Templin (US), Fred L <
> [e-mail]> wrote:
>
> Tom, some rejoinders:
>
>
>
> >Yes, I agree if the packet is fragmented by the network then this is a
> nice feature.
>
> >However, today we already have this from a host perspective property by
> just
>
> >sending "small" packets.
>
>
>
> It can be readily shown that some applications get much greater
> performance by
>
> sending larger packets that trigger fragmentation/reassembly than by
> sending
>
> smaller packets that do not. Multiple order of magnitude performance
> increases
>
> are indeed possible.
>
>
>
> >I'm not sure the savings qualify as significant. 9K MTUs are becoming
> common in data centers
>
> >and the standard TCP/IPv6 header is 80 bytes so that's already less than
> 1% overhead.
>
>
>
> I think 9K is only a starting point, and IP parcels pave the way to much
> larger link MTUs,
>
> possibly even in excess of 64KB. And, doing the math, even for just a 9K
> link sending a
>
> single parcel that contains 6x 1440 octet segments would save 5 * 60 ==
> 300 octets in
>
>
>
> Why would someone put six segments in a parcel if they already have a 9K
> link MTU? Why not just send one segment in 9K?
>
>
>
> comparison with sending 6x  1500 octet packets with 60 octets of IP/TCP
> headers per
>
> packet. For links with larger MTUs, the savings for sending parcels with
> lots of segments
>
> (up to 64) becomes even greater.
>
>
>
> >As I already mentioned, this is addressed by the BiGTCP work (
> https://lwn.net/Articles/884104).
>
> >Sending or receiving multi-megabytes TCP segments in one system call is
> now feasible. Also, it's
>
> >inevitable that NIC vendors will apply this also to be able to offload
> TCP jumbo grams. Given this
>
> >is just software that doesn't require hardware change or on-the-wire
> protocols to change, it's
>
> >immediately deployable with just a softwar change which is a huge benefit
> to datacenter operators.
>
>
>
> As I have said, IP parcels has the same advantage within the host
> system-call (user-space
>
> to kernel-space) context. But, IP parcels goes a step further to provide
> efficient packaging
>
> over-the-wire, whereas the approach you are referring to opens the box
> inside the
>
> kernel and sends individual packets instead of aggregates.
>
>
>
> >All modern NIC HW can deal with offloading a single checksum per packet,
> it's going to be
>
> >a major effort for them to offload multiple checksum like IP parcels
> needs. Without checksum
>
> >offload, this would be a non-starter for a lot of deployments.
>
>
>
> Check the latest spec (now at -12 and likely to stay that way until
> IETF114. Any H/W checksum
>
> that can run over the first segment of a packet should be possible to make
> run over the N-1
>
> additional segments of the same packet (parcel) by applying the very
> familiar Internet
>
> checksum algorithm.
>
>
>
> The algorithm isn't the problem, it's supporting new protocols and
> multiple checksums in a packet in hardware.
>
>
>
>
>
> >I'm not convinced of that. For instance, I'm skeptical that intermediate
> devices trying to reassemble
>
> >packets that aren't addressed to themselves could ever be robust or
> efficient (i.e. complexity, non-work
>
> >conserving resource requirements, security issues with reassembly,
> multi-path that causes latency
>
> >increase, potential DoS vector, etc.). Can you comment on this?
>
>
>
> Perhaps what is confusing this matter is that the intermediate devices
> referred to
>
> here most certainly do not refer to all routers in the path. Instead, what
> is intended
>
> here is an OMNI intermediate device, of which there may be something on
> the order
>
> of 0, 1, or 2 of them on the path between the OMNI source and destination
> even
>
> though there may be many 10’s or even 100’s of ordinary IP routers on the
> path.
>
> And, again, this is not a strict reassembly case – instead, it is an
> opportunistic
>
> “combine if convenient; else forward” swift decision.
>
>
>
> Either you're trivializing reassembly or maybe you're thinking of some new
> method that somehow avoids all the pitfalls and problems we've had with
> reassembly over the years! Consider that many NIC vendors have tried, and
> largely failed, to get any sort of device reassembly widely deployed (e.g.
> IP reassembly, TCP segmentation reassembly, etc.). The reason they failed
> is because they can't give the host stack transparency and control over the
> reassembly process.
>
>
>
> In its nature reassembly can only be done with at least packets. That
> means a device performing reassembly has to receive one packet, hold it,
> and wait for the following packet to perform reassembly. That makes
> reassembly, unlike fragmentation, a non-work conserving process. Many
> issues and policies arise from this. For instance, what happens if a packet
> is held and the following packet is never seen? (usually implies a
> reassembly timer). What happens if a packet is received OOO and is already
> forwarded, but the preceding packet is then received, do we try to
> reassemble that one? (the solution here seems to be to maintain some sort
> of flow state)? What about overlapping fragments and the security issues
> around that?
>
>
>
> IMO, if the WG does pursue this, I believe a lot of the effort will be in
> specifying how reassembly in intermediate nodes works.
>
>
>
> Tom
>
>
>
>
>
> Thanks - Fred
>
>
>
> *From:* Tom Herbert [mailto:[e-mail]]
> *Sent:* Monday, July 11, 2022 1:34 PM
> *To:* Templin (US), Fred L <[e-mail]>
> *Cc:* Richard Li <[e-mail]>; Juan Carlos Zuniga
> (juzuniga) <juzuniga=[e-mail]>; [e-mail]
> *Subject:* [EXTERNAL] Re: [Int-area] Call for WG adoption of
> draft-templin-intarea-parcels-10
>
>
>
> EXT email: be mindful of links/attachments.
>
>
>
>
>
>
>
>
>
> On Mon, Jul 11, 2022 at 12:22 PM Templin (US), Fred L <
> [e-mail]> wrote:
>
> Richard and others, thank you for these comments and for the ensuing
> discussion that
>
> took place over the time I was away on vacation. Strange how the timing
> hit when I
>
> was away from the office and off the grid - I was on a camping trip in
> Canada not far
>
> from where Steve Deering lives although I did not visit him.
>
>
>
> In any event, I was able to push out a new draft version ahead of the
> deadline that
>
> may address some (but likely not all) of your concerns:
>
>
>
> https://datatracker.ietf.org/doc/draft-templin-intarea-parcels/
>
>
>
> The major change is that the draft now talks about interactions with upper
> layer
>
> protocols including TCP and UDP, whereas the previous draft versions were
> silent
>
> regarding upper layer protocol framing.
>
>
>
> To others who have commented, I beg to differ and maintain that IP parcels
> do
>
> represent a significant improvement over the current state of affairs and
> over
>
> just regular IP jumbograms. In particular:
>
>
>
> Hi Fred, some comments in line.
>
>
>
>
>
> 1) IP parcels make it so that the loss unit is a single segment instead of
> the entire
>
> packet/parcel, and loss of a segment often results in retransmission of
> just that
>
> segment instead of the entire packet/parcel.
>
>
>
> Yes, I agree if the packet is fragmented by the network then this is a
> nice feature. However, today we already have this from a host perspective
> property by just sending "small" packets.
>
>
>
>
>
> 2) IP parcels are more efficient than sending a single segment per IP
> packet, since
>
> the parcel includes a single IP header plus single full {TCP,UDP} header
> for possibly
>
> many segments. This can result in significant savings in terms of bits
> over the wire
>
> for omitting unnecessary header bytes.
>
>
>
> I'm not sure the savings qualify as significant. 9K MTUs are becoming
> common in data centers and the standard TCP/IPv6 header is 80 bytes so
> that's already less than 1% overhead.
>
>
>
> Consider the postal service analogy; when
>
> many items can be sent together in a single package/parcel there is a
> large savings
>
> in shippeing and handling costs than when each individual item is shipped
> separately.
>
>
>
> As I already mentioned, this is addressed by the BiGTCP work (
> https://lwn.net/Articles/884104). Sending or receiving multi-megabytes
> TCP segments in one system call is now feasible. Also, it's inevitable that
> NIC vendors will apply this also to be able to offload TCP jumbo grams.
> Given this is just software that doesn't require hardware change or
> on-the-wire protocols to change, it's immediately deployable with just a
> softwar change which is a huge benefit to datacenter operators.
>
>
>
> 3) IP parcels improve large packet integrity by including a separate
> checksum for
>
> each segment instead of a single checksum for the entire packet.
>
>
>
> All modern NIC HW can deal with offloading a single checksum per packet,
> it's going to be a major effort for them to offload multiple checksum like
> IP parcels needs. Without checksum offload, this would be a non-starter for
> a lot of deployments.
>
>
>
> This means that
>
> large parcels (up to a few MB) can be sent in one piece over links with
> sufficiently
>
> large MTU without requiring the link itself to provide strong integrity
> checks over
>
> the entire length of the parcel. This means that link MTUs significantly
> larger than
>
> 9KB are now safely possible.
>
>
>
> 4) IP parcels offer all of the efficiency advantages to upper layers as
> are offered
>
> by GSO/GRO, etc. but also provide benefits 1) through 3) above that are not
>
> offered by GSO/GRO.
>
>
>
> Most of this is doable in GSO/GRO.
>
>
>
>
>
> 5) Plus, the idea is just plain neat. Better packaging is good. More
> efficient
>
> handling is good. Reduced header overhead is good. SAFE larger MTUs are
>
> good. The idea itself is good.
>
>
>
> I'm not convinced of that. For instance, I'm skeptical that intermediate
> devices trying to reassemble packets that aren't addressed to themselves
> could ever be robust or efficient (i.e. complexity, non-work conserving
> resource requirements, security issues with reassembly, multi-path that
> causes latency increase, potential DoS vector, etc.). Can you comment on
> this?
>
>
>
> Tom
>
>
>
>
>
> Fred
>
>
>
> *From:* Int-area [mailto:[e-mail]] *On Behalf Of *Richard
> Li
> *Sent:* Friday, July 01, 2022 3:11 PM
> *To:* Juan Carlos Zuniga (juzuniga) <juzuniga=[e-mail]>
> *Cc:* [e-mail]
> *Subject:* Re: [Int-area] Call for WG adoption of
> draft-templin-intarea-parcels-10
>
>
>
> Chairs and Authors,
>
>
>
> I always like every new idea and effort to improve the Internet
> performance, and thus I have read this draft with a great interest. The
> following are my observations/comments/questions. If they don’t make any
> sense to you, please accept my apology, and disregard them.
>
>
>
> 1.      The text “multiple upper layer protocol segments” is ambiguous.
> It seems that you really mean “multiple segments from ‘the same’ upper
> layer protocol”, doesn’t it? It seems that multiple segments from different
> upper layer protocols are not allowed in your parcel.
>
>
>
> 2.      Is the following a fair statement? All segments in the same
> packet come from the same application identified by the 5-tupe (source
> address, destination address, source port, destination port, protocol
> number).
>
>
>
> 3.      Segment size
>
> You require that their sizes be the same except for the last one. Is this
> required for easy implementation or what? Do you require it for any other
> reasons?
>
>
>
> 4.      TTL issue
>
> You described how parcels are forwarded over the Internetwork, and in
> particular you described what the ingress/egress middlebox does about
> parcels. I understand that the ingress middlebox may break the parcel into
> smaller ones, which may rejoin at the egress middlebox. My question is
> about TTL. As different smaller parcels may traverse along different paths,
> as a result their TTLs may be different when they reach the egress
> middlebox . How does the egress middlebox set up the TTL value? Please
> provide more descriptions.
>
>
>
> 5.      Reordering at the egress middlebox
>
> The parcels would arrive one after another, and therefore the egress
> middlebox would “wait” for a little bit to identify and pick up enough
> parcels/packets for their rejoining and repackaging. A description of the
> egress middlebox behavior would be useful and helpful, in particular I
> would like to know more about the waiting time if any, and how you deal
> with the reordering and loss.
>
>
>
> 6.      IPv4 option
>
> Does IETF still allow to change/add IPv4 option fields? I might be wrong,
> but aren’t they frozen? Also, do commercial routers still care about IPv4
> options?
>
>
>
> 7.      IPv6 option
>
> This draft has defined a hop-by-hop option, it will require every
> intermediate IPv6 router to inspect this option. There have been some
> discussions on the pros/cons about Hop-by-Hop IPv6 Option. Is there any
> feedback from WG 6man?
>
>
>
> 8.      Parcel Path Qualification
>
> This draft has described a method for parcel path qualification probe from
> end to end. It is nice to have it, but it is unreliable simply for the
> following reason: a probe parcel goes along one specific path, and your
> real application parcels may take different paths.
>
>
>
> 9.      Integrity
>
> First paragraph of Section 7. More explanation/elaboration should be
> useful. I might have missed it in previous paragraphs, but if I do, please
> provide a reference to it such as “as described in …”.
>
>
>
> 10.   Implementation Status
>
> In section 10. TSO’s performance gain and Parcel’s gain should be regarded
> as two different things. Since this draft is adding a hop-by-hop option,
> every intermediate router is required to process the hop-by-hop option,
> which will, theoretically speaking, lead to performance downgrade. Of
> course, the whole performance would depend on many other factors, such as
> the total numbers of routing table lookups and number of segments.
>
>
>
> 11.   General observation
>
> This proposal essentially tries to solve a problem caused by MTU. If MTU
> be very big, one would simply put the whole data in a single packet. Since
> MTU is limited, a packet has to be cut into many smaller pieces (segments).
> In the existing specification, when an intermediate router sees a packet
> with its size larger than MTU, the router would be expected to fragment it
> so that the fragments could be forwarded. Here let me call it
> “fragmentation as needed”. In reality, however, some (if not all)
> commercial routers don’t do “fragmentation as needed”, instead of
> fragmenting the packet they simply discard it in order to achieve the
> wire-speed. This draft defines a new way to address the MTU issue: when a
> router sees a packet with its size larger than MTU, the router is asked to
> fragment it in a prescribed way (fragment it into pre-packaged segments).
> If I may, let me call it “fragmentation as prescribed”. Both “fragmentation
> as needed” and “fragmentation as prescribed” would require the support from
> intermediate routers. As the same as fragmentation as needed, fragmentation
> as prescribed may downgrade the performance of intermediate routers. What
> is more, intermediate routers/boxes may perform “rejoining and
> repackaging”, which will adversely impact the performance of the
> intermediate routers/boxes.
>
>
>
>
>
> Best regards,
>
>
>
> Richard
>
>
>
>
>
>
>
> *From:* Int-area <[e-mail]> *On Behalf Of *Juan Carlos
> Zuniga (juzuniga)
> *Sent:* Wednesday, June 22, 2022 12:25 PM
> *To:* [e-mail]
> *Subject:* [Int-area] Call for WG adoption of
> draft-templin-intarea-parcels-10
>
>
>
> Dear IntArea WG,
>
>
>
> We are starting a 2-week call for adoption of the IP-Parcels draft:
>
> https://www.ietf.org/archive/id/draft-templin-intarea-parcels-10.html
> <https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.ietf.org%2Farchive%2Fid%2Fdraft-templin-intarea-parcels-10.html&data=05%7C01%7Crichard.li%40futurewei.com%7C715b5db213134932c70208da5484f702%7C0fee8ff2a3b240189c753a1d5591fedc%7C1%7C1%7C637915227299598680%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000%7C%7C%7C&sdata=w4G5ypaSRv%2FR31%2F%2B857XT2xUqHdEXv90ubD5GGjqBEQ%3D&reserved=0>
>
>
>
> The document has been discussed for some time and it has received multiple
> comments.
>
>
>
> If you have an opinion on whether this document should be adopted by the
> IntArea WG please indicate it on the list by the end of Wednesday July 6th
> .
>
>
>
> Thanks,
>
>
>
> Juan-Carlos & Wassim
>
> (IntArea WG chairs)
>
>
>
> _______________________________________________
> Int-area mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/int-area
>
>
>
> _______________________________________________
>
> Int-area mailing list
>
> [e-mail]
>
> https://www.ietf.org/mailman/listinfo/int-area
>
> _______________________________________________
> Int-area mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/int-area
>