Re: [L2tpext] draft-ietf-l2tpext-l2tphc-04.txt

W. Mark Townsley writes:
> 
> Please see inline.
> 
> Thanks,
> 
> - Mark
> 
> Skip Booth wrote:
> > 
> > Thomas Narten writes:
> > > The WG has asked to publish this document as a PS, but I have some
> > > questions I'd like to get WG input on first.
> > >
> > > General concerns
> > >
> > > 1) How much use would L2TPHC as described in this document actually
> > >    see in practice. It seems to be of limited applicability, given the
> > >    assumptions being made. Specifically, the document says:
> > >
> > >    If several simplifying assumptions may be met, it is possible to
> > >    reduce the size of the L2TP encapsulation:
> > >
> > >       - The tunnel will not operate through a NAT interface
> > >       - The tunnel uses a single IP address for the life of the tunnel
> > >       - The tunnel's host uses only one public IP network interface
> > >       - There will be only one tunnel between the LAC and the LNS
> > >       - There might be only one session within a tunnel
> > >       - There might be only one protocol active on that session
> > >       - Alignment is not required
> > >       - Packet length is preserved by the IP header
> > >
> > >    Where will the compression scheme described in this document be
> > >    deployed, given its restricted applicability?
> 
> I think l2tphc does a nice job of identifying all of the reasons why you may NOT
> want to run with most of your l2tp and ppp header missing. And, yes, there are a
> lot of very good reasons not to do this. The efficiency in number of bytes on
> the wire has to be of such overriding concern that none of these items remain
> significant. To not point these out would be a very bad thing.
> 
> > >
> > > 2) The protocol is firewall hostile, as it's carrying a dynamic
> > >    payload that sits directly on top of IP. There is nothing above the
> > >    IP protocol to filter on, since it may carry an arbitrary raw PPP
> > >    frame. Will firewalls in fact let such traffic through? Should
> > >    they?  How will this impact deployability?
> 
> Agreed. Daniel identified that the intended application does not require NAT
> friendliness. If l2tphc doesn't underscore this loudly enough now, it should.
> 
> > >
> > > 3) The document requires assignment of an IP protocol number, a
> > >    limited resource. If this technique has only limited applicability,
> > >    it would not seem to justify assignment of an IP protocol
> > >    number. I'd like to better understand  whether an IP port number
> > >    can be justified for this application/usage
> 
> We could always resurrect the negotiation of the protocol ID in the l2tphc
> setup, as was done in the early versions of the draft. It's not very NAT
> friendly, but as Daniel pointed out, the known intended application of l2tphc
> doesn't utilize a NAT, and as Thomas pointed out, l2tphc isn't very NAT friendly
> anyway (and neither is L2TPv3's IP mode, FWIW). I think I would rather see
> L2TPv3 used here instead (see discussion later in this email), but it is still a
> possibility.
> 
> > >
> > > 4) Is this use of compression compatable with IPsec? I.e., do existing
> > >    IPsec implementations support protocols other than TCP/UDP  when
> > >    run over IP? In this case, L2TP is being run directly above IP.
> 
> When you stop running over UDP, you lose a very handy selector for demuxing SAs
> between like IP address pairs. However, you can still setup an SA based on your
> IP address pair alone for the L2TPHC protocol type. Note that this is really no
> more limited than, say, running GRE on top of IP and protecting it with IPsec.
> 
> > >
> > > 5) How does this compare to other existing, more general compression
> > >    techniques. Sure, you can always squeeze a few more bytes by having
> > >    something very application-specific, but in general this is not
> > >    viewed as a good thing, especially when (as in this case) an IP
> > >    protocol number is needed and the usage applies to only limited
> > >    configurations where L2TP is in use. How does this compare with,
> > >    say, RFC 3095 compression.
> 
> I think Andy's response to this was that RFC3095 didn't exist at the time l2tphc
> was written, or even when it first entered last call. If the folks that intend
> to use l2tphc (UTRAN?) can use RFC 3095 to achieve the same gains in a more
> general manner, great. Personally, I haven't even tried to investigate that and
> would be happy to hear why it is or is not applicable.
> 
> > >
> > > Note: I also don't quite agree with "5. Efficiency Considerations".
> > >
> > > >    Some rough calculations will illustrate the environments in which
> > > >    L2TPHC may be beneficial.  Overhead as a percentage of the carried
> > > >    traffic will be calculated for a typical packet size involved in bulk
> > > >    data transfer (700 bytes), and the canonical 64-byte ``small IP
> > > >    packet''.  Percentages will be rounded to the nearest whole number.
> > > >    Overhead is tallied for an IP header of 20 bytes, a UDP header of 8
> > > >    bytes, an L2TP header of 8 bytes, and a PPP encapsulation of 4
> > > >    bytes.
> > >
> > > >
> > > >    The worst case is a 64-byte packet carried within a UDP L2TP header.
> > > >    The 64 bytes of payload is carried by an overall header of 40 bytes,
> > > >    resulting in an overhead of 63%.  With the larger payload size of 700
> > > >    bytes, the header is amortized over many more bytes, reducing the
> > > >    overhead to 6%.
> > >
> > > Seems to me that the overhead being listed is misleading. If the
> > > original payload is 64 bytes (this includes a fully formed TCP/IP/PPP
> > > packet with its own IP header), and one then encapsulates it within
> > > L2TP, you are adding 36 bytes of overhead by using l2tp. That's an
> > > overhead of 36/(68+36) = 34% (when compared to the no l2tp case).
> 
> Regarding overhead calculations, I believe Andy already said that he would look
> into these as long as there still was support for l2tphc in general. I
> understand his desire to not spend any time on the draft at this point, as it
> has been tugged back and forth for a long while now. Thanks for a very thorough
> job of digging into the details here, Thomas, if you are right then we surely
> should have caught this a lot time ago. 
> 
> > >
> > > I.e., counting the original IP/PP headers as part of "overhead" is
> > > misleading because they would get sent anyway, and those are not being
> > > compressed away by this technique (at least as I understand things).
> > >
> > > >    With L2TPHC, the UDP header is absent, the L2TPHC header is 0 bytes
> > > >    for the most compact case, and the 4 bytes of PPP encapsulation have
> > > >    been deleted.  Overall size is thus 20 bytes of IP header.  The small
> > > >    packet now suffers an overhead of only 31%, and the larger packet 3%.
> > >
> > > So now your overhead  is more like 20/(64+20) = 23%.
> > >
> > > So yes, there is a reduction in overhead, but its not 63% to 31%, its
> > > more like 34% to 23%.
> > 
> > With L2TPv3 coming along the difference is even less.  Since L2TPv3
> > proposes a 4 byte L2TP header in the simplest case and it runs directly on
> > top of IP, the total header size is 24 bytes minimum.  So now we are down
> > to a 4 byte difference in header sizes.  IMO, that's not worth proposing a
> > new RFC for and allocating a IP protocol number for.
> 
> Good point, Skip. 
> 
> However, one could still utilize the l2tphc mechanisms to whittle down
> the PPP framing if necessary, so it is a little more than just the 4
> bytes (PPP's mechanisms can get you down to one byte, but l2tphc gets you
> to 0 which might be significant if you are worried about an aligned inner
> header as well as a minimal header).

Negotiating away the PPP protocol header should be a PPP option not a L2TP
HC feature.  IMO this shouldn't be a part of this draft to start with.

>  Add to this the fact that the PPP over L2TPv3 header has its own 4 byte
> header for sequencing and offset (I don't think we have made it clear
> that it is not always present, which something l2tphc-like built into the
> PPP spec would have to mandate) and the bytes could start to add up.

Since sequencing and offset was never a required part of RFC2661 I think
it's fair to assume that these are not part of the PPP over L2TPv3 draft.

So assuming the negotiation of the PPP protocol off is a separate PPP draft
and we don't need sequencing/offset information for PPPoL2TPv3, we are
comparing a 4 byte L2TPv3 header to a 1 byte L2TPHC header.  Is a new RFC
just to save 3 bytes really worth it?  I'd much rather see the investment
in L2TPv3 which is something we really need in order to solve a much larger
set of problems.

-Skip

> 
> If there are existing interoperable implementations, or the need to reduce the
> l2tp header by the 5-9 bytes that l2tphc provides you on top of the new IP encap
> in L2TPv3 for some applications is deemed highly significant (the UTRAN folks
> are as good a candidate as anyone to argue this case), then I think we should go
> ahead and publish the RFC. 
> 
> However, if folks can live with the 4 byte IP encapsulation defined in
> draft-ietf-l2tpext-l2tp-base-02.txt, then we could simply (1) add something to
> the PPP over L2TPv3 spec that ensures the PPP control word may not be present,
> and (2) negotiate away the PPP framing ala l2tphc now. This would be something
> that could be folded into the PPP over L2TPv3 spec, or companion spec (perhaps
> easily negotiated via L2TPv3's new PW-Type AVP as a separate "compressed PPP"
> PW-Type). This would at least have the advantage of fewer L2TP encapsulations to
> support overall, and could fit into the L2TPv3 PW model well.
> 
> - Mark
> 
> > 
> > My 2c.
> > 
> > -Skip
> > 
> > >
> > > Thomas
> > >
> > > _______________________________________________
> > > L2tpext mailing list
> > > L2tpext@ietf.org
> > > https://www1.ietf.org/mailman/listinfo/l2tpext
> > 
> > _______________________________________________
> > L2tpext mailing list
> > L2tpext@ietf.org
> > https://www1.ietf.org/mailman/listinfo/l2tpext

_______________________________________________
L2tpext mailing list
L2tpext@ietf.org
https://www1.ietf.org/mailman/listinfo/l2tpext