Re: [Int-area] IP parcels

[Tom Herbert <[e-mail]>]

On Tue, Dec 21, 2021 at 6:24 AM Templin (US), Fred L
<[e-mail]> wrote:
>
> Tom, reading your message makes me think you have not read my drafts. The
>
> answers to the perceived issues you are raising are all there. I do not see anything
>
> new in what you are saying to make me believe otherwise.
>
>

Fred,

I did read your draft. I might be misunderstanding it. Here are points
I don't think I understand, f these are in the draft plead reference
the precise section:

- A clear explanation why GSO/GRO are not sufficient to solve the
problem. The draft highlights these to show the advantages of
sending/receiving large data units, but it's not clear to me why a
change to the protocol is required to get the same or somehow better
effect. As a side note, it should be pointed out that GSO/GRO and
similar mechanisms are opportunistic optimizations. For instance, the
TCP congestion window and receive window have to be large enough for
GSO to be effective for sending on a connection. As an anecdote, there
was an incident early on at YouTube where they extensively used GSO
for serving video which under normal circumstances is a great savings
in CPU utilization. But one day there was a hiccup on the Internet
that caused all the connections to go to slow start. So now instead of
sending 64K at a time the servers were sending two segments at a time
for all the connections (this was before the work to raise initcwnd);
so now instead of servers running at 50% CPU, they needed 150% CPU and
so were dropping a lot of packets, and recovering took quite a bit of
time making unhappy customers (moral of this story: always provision
your servers to handle the worst case scenario where opportunistic
optimizations become ineffective!)

- What is the exact algorithm for reassembly of parcels? Searching the
document for reassemble only comes up with "when the OAL source or
final destination receives the fragments or whole parcels, it
reassembles if necessary"

- What are the ramifications of middleboxes performing reassembly on
behalf of a host. The document says "then rejoined into one or more
parcels at a last-hop middlebox to be forwarded to the final
destination". I'm not sure what a "last-hop middlebox" means in a
normative context, but this does appear to be an intermediate network
node which would seem to be susceptible to the issues of stateful
intermediate network nodes that I previously raised.

- Is in order delivery of segments within a parcel maintained. The
draft states "While not desirable, reordering of segments within
parcels and individual segment loss are possible.  But, what matters
is that the number of parcels delivered to the final destination
should be kept to a minimum, and that loss or receipt of individual
segments (and not parcel size) determines the retransmission unit".
Why is it so critical to keep the number of parcels delivered to the
final destination to be kept at minimum? As I mentioned, hosts are
already used to dealing with reassembly, it seems like the best method
is still to send packets at path MTU which is what TCP is doing with
PMTUD.

- How are IP parcels substantially different from fragmentation?  Is
the idea that individual segments in an IP parcel can be lost without
losing the whole parcel? Is the idea that parcels can make up a >64K
super packet? What if a segment in a parcel is greater than an MTU in
the path, is an intermediate node breaking up a parcel expected to
fragment the segment, or send a PTB?

Tom



>
> Thanks - Fred
>
>
>
> From: Tom Herbert [mailto:[e-mail]]
> Sent: Monday, December 20, 2021 4:14 PM
> To: Templin (US), Fred L <[e-mail]>
> Cc: [e-mail]; [e-mail]
> Subject: Re: IP parcels
>
>
>
>
>
> On Mon, Dec 20, 2021 at 3:11 PM Templin (US), Fred L <[e-mail]> wrote:
>
> Tom, in modern reassembly it is not going to wait for the MSL for all fragments
>
> to arrive anymore; either they all get there after a very small inter-fragment
>
> delay, or you send an immediate FRAGREP and possibly also a PTB soft error
>
> then quickly declare the reassembly dead if that doesn’t help. And, you make
>
> sure to inspect IDs of received fragments before admitting them into the
>
> reassembly cache so you don’t end up caching garbage that will just have to
>
> be discarded later.
>
>
>
> Fred,
>
>
>
> It doesn't matter in the sense that reassembly is a non-working conserving mechanism. In order to perform reassembly packet fragments need to be held which means memory will be consumed and since memory is a finite resource it needs to be managed.  Managing memory means that some policy is needed when to time out a reassembly or which fragment train to discard under memory pressure. A network that implements some arbitrary policy can cause problems on unsuspecting hosts. For instance, there's mechanisms for hosts to try to guess what the timeout is in a NAT box and send a keepalive packet before an idle NAT state is evicted. So this is just a guess that may or may not be right, and in fact there might not even be a NAT in the path in which case the host is just wasting energy sending keepalives. Also, the second we introduce a new exhaustible resource in the path that becomes yet another denial of service vector (consider the case that an attacker spoofs a whole bunch of IP parcels).
>
>
>
> Unless the network can coordinate very specifically with the host about what it's doing on behalf of the host stack, I think it's much better for the network to just focus or forward packets without delay and let the host handle the details of receive processing, reassembly, security, etc.
>
>
>
> Tom
>
>
>
>
>
> Fred
>
>
>
> From: Tom Herbert [mailto:[e-mail]]
> Sent: Monday, December 20, 2021 1:06 PM
> To: Templin (US), Fred L <[e-mail]>
> Cc: [e-mail]; [e-mail]
> Subject: Re: IP parcels
>
>
>
>
>
> On Mon, Dec 20, 2021 at 12:03 PM Templin (US), Fred L <[e-mail]> wrote:
>
> Tom, sorry I will try to use my words more carefully; I am using GSO/GRO also for
>
> a UDP-based transport protocol – not QUIC but something similar. I like GSO/GRO
>
> very much; I am glad the service is available and I want to see it continue. But, my
>
> understanding of the services is that they leverage the IP ID field in whole IPv4
>
> packets that are not eligible for fragmentation and those are limitations I am
>
> seeking to improve on.
>
>
>
> I want to enable a facility similar to GSO/GRO that works for both IPv4 and IPv6
>
> packets and allows for lower layers to fragment if necessary. And, I want to use
>
> a well-behaved 32-bit IPv6 ID instead of the 16-bit IPv4 one where the use is not
>
> well defined when DF=1.
>
>
>
> There has been a lot of work in this area. For instance, you might want to take a look at https://www.youtube.com/watch?v=ccUeG1dAhbw
>
>
>
> About reassembly, that would only happen on the end systems themselves or on
>
> a very capable device that is very close to the end systems; I would not want for
>
> a high-speed core router to have to reassemble.
>
>
>
> Even so, an intermediate device close to the end system still has to provide service to more than one host. Reassembly requires memory to store fragments. A host would need enough memory to service all of its own flows, but an intermediate node would need number of hosts it serves times that amount of memory to perform reassembly.  This is a fundamental scaling problem of stateful services in the network, inevitably the network nodes cannot scale to the number of users or flows that require service. In the best case scenario, when resources are not available the network won't attempt the stateful operation and will just forward the packet unimpeded (which is fine because host will never rely on this class of optimization). In the worse case scenario, the network will take a detrimental action such as forcibly breaking a connection (e.g. this is what can happen when a NAT evicts a TCP connection because it has run out of memory). IMO, maintaining state in the network is a bad, albeit unfortunately prevalent, idea.
>
>
>
> Tom
>
>
>
>
>
> Again, GSO/GRO is nice work and much respect is due to those who made it possible.
>
>
>
> Fred
>
>
>
> From: Tom Herbert [mailto:[e-mail]]
> Sent: Monday, December 20, 2021 9:20 AM
> To: Templin (US), Fred L <[e-mail]>
> Cc: [e-mail]; [e-mail]
> Subject: Re: [Int-area] [EXTERNAL] Re: IP parcels
>
>
>
> The world is not just TCP anymore. QUIC and other UDP-based transports have already
>
> shown performance increases using facilities like GSO/GRO which are essentially a short
>
> term and non-standard implementation of what parcels promise to do in the long term.
>
>
>
> Fred,
>
>
>
> Can you explain why GSO/GRO aren't sufficient and are only short term solutions? We've been using these for almost twenty years now with good effect. These are widely deployed with TCP, TSO works well to offload transmit, LRO is defined and is in much better shape to offload RX now that programmable devices are emerging. For TCP it's hard to see how IP parcels would help significantly, but even for UDP we now have UDP GSO, sendmmsg, and recvmmsg that mitigate the cost of system calls and interrupts to which the draft refers. The reason these aren't standards in IETF is because they're implementation techniques and not protocol (although I will point out that GSO/GRO/sendmmsg/recvmmsg are in all Linux devices so that effectively makes it a de facto implementation standard).
>
>
>
> I am also concerned about the idea that intermediate devices would perform reassembly. This has a whole bunch of implications like middleboxes are no longer work conserving and seems to have the implicit requirement that it has to be in the path of every packet in a parcel (i.e. even in the case of the last hop performing reassembly. Also, as simply a matter of resources and capabilities, hosts are in a much better position to perform tasks like reassembly. I don't readily see that having intermediate devices perform reassembly would be a win for hosts, and even if it were, host implementations still would need the capability to perform reassembly themselves since they will never rely on the network to always do it for them.
>
>
>
> Tom
>
>
>
>
>
> Thanks - Fred
>
>
>
> From: [e-mail] [mailto:[e-mail]]
> Sent: Sunday, December 19, 2021 11:53 AM
> To: Templin (US), Fred L <[e-mail]>
> Cc: [e-mail]; Wes Eddy <[e-mail]>
> Subject: Re: [Int-area] IP parcels
>
>
>
> Hi, Fred (et al.),
>
>
>
> On Dec 19, 2021, at 10:21 AM, Templin (US), Fred L <[e-mail]> wrote:
>
>
>
> Joe, your insistence on using html makes it impossible to respond to all of your points inline
>
> which is the reason for my top-posts.
>
>
>
> I use MacOS mail, IOS mail, and Thunderbird on Windows, all using default configurations, FWIW. I appear to be able to post inside everyone else’s responses. I don’t know if the IETF’s mailers are munging formats, though.
>
>
>
> I’ve made my position clear. However:
>
>
>
> - You still haven’t shown any evidence that end systems need to do all this extra work so they can somehow run faster, nor that this will be noticeably faster than large (i.e., 20-60KB) IPv4 packets.
>
>
>
> - You still haven’t shown any reason why this is feasible; in fact, below you add the idea of on-path fragmentation, which is largely deprecated because fragments won’t traverse tunnels (in your case, notably for single chunks larger than 64KB). Nevermind that the fragmentation is both expensive and slow-path at routers.
>
>
>
> - You have claimed that both routers and transports will somehow adopt this when we can’t even get reasonably large MTUs that already fit within IPv4 across heterogeneous enterprises.
>
>
>
> IPv4 is over; even if you don’t think so, any way forward with larger packets starts with:
>
>                a) getting ~64KB IP packets across the net
>
>                b) after (a), prove that >64KB are needed based on the IPv6 jumbo approach
>
>
>
> Any way forward with a lot of small packets inside one large one (where both chunks and total length are less than 64K) starts by proving there’s a need and it fixing how TCP interacts with its inherent burstiness and loss correlation.
>
>
>
> Only THEN will this issue be worth more discussion.
>
>
>
> Joe
>
>
>
>
>
> Parcels that contain a single segment whether 64K or considerably less are still sent as
>
> (singleton) parcels and not ordinary packets. That way, nodes in the network can know
>
> that it is OK to encapsulate and fragment since by asserting its interest in receiving parcels
>
> the destination has also subscribed to being able to reassemble up to a full 64K.
>
>
>
> Parcels do not set (Payload Length / Total Length) to 0; they set it to the length of the
>
> first element of the parcel (which is also the same length of each non-final element of
>
> the parcel). What happens then is that network equipment will see a unit with an L3
>
> length that may be considerably shorter than the L2 length. You are right that legacy
>
> routers might not like this (or, they might truncate the packet according to L3 length),
>
> and so for paths that might traverse legacy routers the first-hop node that recognizes
>
> parcels instead encapsulates the parcel in an IPv4 or IPv6 header then performs (source)
>
> fragmentation if necessary. These IP fragments will then travel through legacy routers
>
> just fine.
>
>
>
> About RFC793bis, you and Wes Eddy know far more about its status than I do; I only
>
> noted that this is something with TCP implications and so made mention of it in case
>
> there is still room for a few more engine tweaks while the hood is still open.
>
>
>
> About IPv4, I am currently running IPv4 edge networks with IPv4-in-IPv6 tunnel endpoints
>
> connected to an IPv6 transit network and it works really good. End systems get to use
>
> smaller addresses and smaller headers, and they can talk to remote correspondents using
>
> IPv4 as if they were all on the same IPv4 network. So yes, I think we might still want to
>
> consider IPv4 for edge networks like that.
>
>
>
> About getting 64K packets across, only the edge networks or end systems see them as
>
> large packets; in the core thy are typically broken up into something much smaller by
>
> ingress nodes that apply segmentation/fragmentation. We don’t need the core to move
>
> to jumbo links; we only need that at the edges. ATM taught us that.
>
>
>
> About our “nail”, end systems get to see larger packets/parcels and get to take advantage
>
> of the reduced interrupts and system call overhead they provide. That is what makes it
>
> worthwhile.
>
>
>
> Fred
>
>
>
> From: [e-mail] [mailto:[e-mail]]
> Sent: Saturday, December 18, 2021 8:13 PM
> To: Templin (US), Fred L <[e-mail]>
> Cc: [e-mail]; Wes Eddy <[e-mail]>
> Subject: Re: [Int-area] IP parcels
>
>
>
> HI, Fred,
>
>
>
> If you have one segment that’s less than 64K, you don’t need the parcel option at all.
>
>
>
> If you have something longer than 64K, either as a single segment or multiple smaller segments, by setting total length to 0, you end up being dropped by legacy routers, which either ignore options they don’t understand or drop packets with options they don’t support.
>
>
>
> RFC793bis does talk about IPv6 jumbos, but this new work is out of scope for RFC793bis - furthermore, it’s too late. It has passed WGLC, IETF LC, and is currently in IESG review for publication.
>
>
>
> You also haven’t addressed why the IETF should be taking up this *new* work for IPv4, which I thought also had been considered ineligible.
>
>
>
> But overall, again, what’s the point? We can’t even get 64K IP packets through the Internet; making them larger doesn’t make that easier or more likely. Such large sizes are of diminishing benefit; routers already forward at 40Gbps per link for minimal packets and end systems have other problems that this exacerbates.
>
>
>
> This seems a lot like a huge hammer in search of a nail. Where’s the nail?
>
>
>
> Joe
>
>
>
> —
>
> Joe Touch, temporal epistemologist
>
> www.strayalpha.com
>
>
>
> On Dec 18, 2021, at 7:18 PM, Templin (US), Fred L <[e-mail]> wrote:
>
>
>
> Joe, I never said that I wanted to restrict this to small transport segments; in fact, I want
>
> just the opposite – I want large segments. A perfectly legal parcel is one which includes 1
>
> ~64KB segment - another legal parcel is one which includes 64 of them! If you want bigger
>
> segments than that, then true jumbos are necessary and this spec does not preclude that.
>
>
>
> About RFC793(bis), I see there is a section on Jumbos and IP parcels is (sort of) an application
>
> of Jumbos.
>
>
>
> Fred
>
>
>
> From: [e-mail] [mailto:[e-mail]]
> Sent: Saturday, December 18, 2021 4:57 PM
> To: Templin (US), Fred L <[e-mail]>
> Cc: [e-mail]; Wes Eddy <[e-mail]>
> Subject: [EXTERNAL] Re: [Int-area] IP parcels
>
>
>
> EXT email: be mindful of links/attachments.
>
>
>
>
> Hi, Fred,
>
>
>
> Regarding 793bis, new ideas are out of scope. It’s supposed to be a roll-in of existing items only.
>
>
>
> Nevermind the problems below, which “TCP will find a way” doesn’t magically fix.
>
>
>
> The problem is this:
>
> - end systems need to send larger transport segments (not just IP segments)
>
> - if they can do that, they should, filling up to the largest IP payload
>
>
>
> Having an IP packet have the opportunity to include lots of small transport packets assumes transport packets either work better or faster when they’re small. It’s the opposite.
>
>
>
> Joe
>
>
>
> —
>
> Joe Touch, temporal epistemologist
>
> www.strayalpha.com
>
>
>
> On Dec 18, 2021, at 4:42 PM, Templin (US), Fred L <[e-mail]> wrote:
>
>
>
> Joe, TCP will find a way to adapt – it always has. I also see that TCP is currently undergoing
>
> a second edition revision so the timing seems right to consider IP parcels in the analysis.
>
> I am Cc’ing the second edition editor for his information – Wesley, please consider this
>
> new concept called “IP Parcels” as it relates to your document.
>
>
>
> Here is the latest draft version – it expands on the “Motivation” section and adds a number
>
> of important feature such as a new “Parcels Permitted” TCP option:
>
>
>
> https://datatracker.ietf.org/doc/draft-templin-intarea-parcels/
>
>
>
> Fred
>
>
>
> From: [e-mail] [mailto:[e-mail]]
> Sent: Friday, December 17, 2021 6:01 PM
> To: Templin (US), Fred L <[e-mail]>
> Cc: [e-mail]
> Subject: Re: [Int-area] IP parcels
>
>
>
> Hi, Fred,
>
>
>
> I’m first concerned at the use of an IP option at all, due to the problems with *any* options forcing processing to slow-path.
>
>
>
> From TCP’s viewpoint, it seems like you’ve just created a nightmare for SACK and ECN, basically because you will encourage drops of large bursts of packets.
>
>
>
> This will also increase the bustiness of TCP, i.e., rather than letting the ACKs support pacing.
>
>
>
> Any part of the system that currently coalesces TCP packets is likely to generate errors here, because they might see only the first TCP segment.
>
>
>
> However, AFAICT the most significant consideration is that  the issue with per-packet performance is at the TCP and UDP layers, not as much at the IP layer.
>
>
>
> So what problem is this trying to solve?
>
>
>
> Joe
>
> —
>
> Joe Touch, temporal epistemologist
>
> www.strayalpha.com
>
>
>
> On Dec 17, 2021, at 5:06 PM, Templin (US), Fred L <[e-mail]> wrote:
>
>
>
> Here's one that should help with shipping, just in time for Christmas. Thanks
> to everyone for the past and future list exchanges.
>
> Fred
>
> -----Original Message-----
> From: I-D-Announce [mailto:[e-mail]] On Behalf Of [e-mail]
> Sent: Friday, December 17, 2021 5:00 PM
> To: [e-mail]
> Subject: I-D Action: draft-templin-intarea-parcels-00.txt
>
>
> A New Internet-Draft is available from the on-line Internet-Drafts directories.
>
>
>        Title           : IP Parcels
>        Author          : Fred L. Templin
>                Filename        : draft-templin-intarea-parcels-00.txt
>                Pages           : 8
>                Date            : 2021-12-17
>
> Abstract:
>   IP packets (both IPv4 and IPv6) are understood to contain a unit of
>   data which becomes the retransmission unit in case of loss.  Upper
>   layer protocols such as the Transmission Control Protocol (TCP)
>   prepare data units known as "segments", with traditional arrangements
>   including a single segment per packet.  This document presents a new
>   construct known as the "IP Parcel" which permits a single packet to
>   carry multiple segments.  The parcel can be opened at middleboxes on
>   the path with the included segments broken out into individual
>   packets, then rejoined into one or more repackaged parcels to be
>   forwarded further toward the final destination.  Reordering of
>   segments within parcels is unimportant; what matters is that the
>   number of parcels delivered to the final destination should be kept
>   to a minimum, and that loss or receipt of individual segments (and
>   not parcel size) determines the retransmission unit.
>
>
> The IETF datatracker status page for this draft is:
> https://datatracker.ietf.org/doc/draft-templin-intarea-parcels/
>
> There is also an htmlized version available at:
> https://datatracker.ietf.org/doc/html/draft-templin-intarea-parcels-00
>
>
> Internet-Drafts are also available by rsync at rsync.ietf.org::internet-drafts
>
>
> _______________________________________________
> I-D-Announce mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/i-d-announce
> Internet-Draft directories: http://www.ietf.org/shadow.html
> or ftp://ftp.ietf.org/ietf/1shadow-sites.txt
>
> _______________________________________________
> 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