Tom, IP parcels have a very significant difference from the GSO/GRO and others
you mentioned
in that IP parcels allow a *single* packet to contain *multiple* upper layer
protocol segments;
in all of the other schemes you cited, it is always a single ULP segment per
packet. This alone
demonstrates that IP parcels at the very least provides a significant savings
in terms of reduced
packet headers, since only a single copy of the {TCP,UDP}/IP headers appears
and not multiple.
The maximum IP parcel size is also not constrained by the underlying network
path MTU the
same way that the maximum GSO/GRO packet size is. So, even if the path MTU is
only 1500
IP parcels up to 64KB and even larger can be sent over an OMNI interface
configured over
the path. If you did that with GSO, then the packets would arrive at the
destination
fragmented and as you know in linux GRO cannot apply UDP/IP reassembly to
packets
that have already undergone fragmentation at the sub-IP layer. Yes, you can
linearize
the packets but the second you do that any GRO performance gains are lost.
You mentioned data centers going to 9KB, and while that is good it is still way
smaller
than what we should be aiming for. IP parcels will encourage links with much
larger
MTUs - 64KB is just a starting point, and going much larger into multiple MBs
can
be a near-term goal. Yes, IP parcels can take full advantage of 9KB MTUs right
away
and still be better than the other schemes because larger MTUs at the sub-IP
layer
result in less IPv6 fragmentation and associated savings in sub-IP layer
encapsulation
overhead.
IP parcels can be thought of as a gateway to larger MTUs in the Internet without
having to compromise integrity and/or without having to retransmit lots of big
blocks of data if only just one or a couple of bits were damaged in transit. The
IP parcels philosophy is to accept as much good data as possible while asking
for retransmission only of errored data that cannot be locally repaired. This
will be good for a vast array of bulk-transfer Internetworking applications, not
only within the local data center but also across the wide area using OMNI.
I could go on, but I won't for now. I have done the work, and I have shown the
work. The community now needs to apply a check-mark to acknowledge.
Fred
> -----Original Message-----
> From: Tom Herbert <[email protected]>
> Sent: Wednesday, November 09, 2022 9:22 AM
> To: Templin (US), Fred L <[email protected]>
> Cc: Richard Li <[email protected]>; [email protected]
> Subject: [EXTERNAL] Re: [Int-area] About draft-templin-intarea-parcels
>
> EXT email: be mindful of links/attachments.
>
>
>
> On Wed, Nov 9, 2022 at 7:47 AM Templin (US), Fred L
> <[email protected]> wrote:
> >
> > Richard, thank you for your message. The intarea community must understand
> > that
> >
> > the live IP Parcels presentation given today was only a “roadmap” to a
> > proper
> >
> > presentation which could not be given due to time constraints. The charts
> > shown
> >
> > during the live presentation were skipped over quickly, but they provide
> > full
> >
> > detail and are permanently available here:
> >
> >
> >
> >
> > https://datatracker.ietf.org/meeting/115/materials/slides-115-intarea-ip-parcels
> >
> >
> >
> > Running code is also now permanently available here:
> >
> >
> >
> > https://github.com/fltemplin/ip-parcels
> >
> >
> >
> > and provides clear evidence that IP parcels provide an appreciable
> > performance
> >
> > gain which cannot be ignored any longer.
> >
> >
> >
> > IP parcels are good for the Internet, and the presentation charts and
> > running code
> >
> > provide clear evidence. It is time to adopt IP parcels.
>
> Fred,
>
> Thanks for the data and implementation, but I'm still not convinced
> that IP parcels should be adopted. Your data seems to show that when
> the networking stack processes large segments performance increases
> (fewer packets to process in the data path is a win). We've known this
> for a long time and that's why stacks commonly implement various
> segmentation techniques like GSO/TSO, GRO/LRO, UFO, USO, and more
> recently BigTCP. Also, within the data center, 9K MTUs are becoming
> common place which is even better than segmentation with 1500 byte
> MTU. The major difference between these techniques and IP parcels is
> that the segmentation techniques don't require any new protocol or
> changes to an existing protocol, whereas IP parcels requires protocol
> changes. So in order to justify IP parcels adoption, not just in IETF
> but also upstreaming into Linux, I think you'll want to show that it
> has significant benefits over the existing segmentation techniques to
> justify the complexities and cost of a new protocol.
>
> Tom
>
> >
> >
> >
> > Fred
> >
> >
> >
> > From: Int-area <[email protected]> On Behalf Of Richard Li
> > Sent: Wednesday, November 09, 2022 6:12 AM
> > To: [email protected]
> > Subject: [Int-area] About draft-templin-intarea-parcels
> >
> >
> >
> > Hi Chairs and All,
> >
> >
> >
> > At today’s intarea meeting, the chair asked the participants if anybody has
> > an interest in this draft or not. If nobody is interested, this draft
> will be closed, and if anyone is interested, he/she is asked to voice it on
> the mailing list.
> >
> >
> >
> > As a follow up, I am expressing my interest in this draft, and would like
> > to see this draft open and let it go on. A few months ago, I asked
> its authors several questions, and the authors answered and clarified them. I
> do see good values for some use cases, especially for those in
> broadband access and jumbo frames being used on the links. It seems to me
> that this draft points to a useful direction, some rooms are
> remaining for expansion though.
> >
> >
> >
> > Thanks,
> >
> >
> >
> > Richard
> >
> >
> >
> > _______________________________________________
> > Int-area mailing list
> > [email protected]
> > https://www.ietf.org/mailman/listinfo/int-area
_______________________________________________
Int-area mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/int-area
