Re: [trill] Tsvart early review of draft-ietf-trill-over-ip-10

[Donald Eastlake <[e-mail]>]

Hi Magnus,

Thanks for the quick response.

On Fri, Feb 2, 2018 at 4:29 AM, Magnus Westerlund <
[e-mail]> wrote:

> Hi Donald,
>
> I have reviewed the changes and my comments. Please see inline for my
> feedback. I will remove text not necessary to provide context. It is almost
> ready now. There are two areas where I think there needs to be some further
> improvements, DSCPs and TCP encapsulation. But the needed changes are
> relatively straightforward.
>
> Den 2018-01-31 kl. 06:25, skrev Donald Eastlake:
>
> Hi Magnus,
>
> It has been a while but the just posted version -12 is intended to
> resolve your comments except those related to middle boxes. (The TRILL
> WG has decided middle boxes will be out of scope for this draft.)
>
>
>
> On Thu, Jun 29, 2017 at 10:17 PM, Donald Eastlake <[e-mail]> <[e-mail]> wrote:
>
> On Tue, Jun 27, 2017 at 1:13 PM, Magnus Westerlund<[e-mail]> <[e-mail]> wrote:
>
> Den 2017-06-26 kl. 02:07, skrev Donald Eastlake:
>
> On Thu, Jun 15, 2017 at 1:32 PM, Magnus Westerlund<[e-mail]> <[e-mail]> wrote:
>
>
> Diffserv usage
> --------------
>
> Section 4.3:
>
>
> The new text from -12:
>
>    The default TRILL priority and DEI to DSCP mapping, which may be
>    configured per TRILL over IP port, is an follows. Note that the DEI
>    value does not affect the default mapping and, to provide a
>    potentially lower priority service than the default priority 0,
>    priority 1 is considered lower priority than 0. So the priority
>    sequence from lower to higher priority is 1, 0, 2, 3, 4, 5, 6, 7, as
>    it is in [802.1Q].
>
>       TRILL Priority  DEI  DSCP Field (Binary/decimal)
>       --------------  ---  -----------------------------
>                   0   0/1  001000 / 8
>                   1   0/1  000010 / 0
>                   2   0/1  010000 / 16
>                   3   0/1  011000 / 24
>                   4   0/1  100000 / 32
>                   5   0/1  101000 / 40
>                   6   0/1  110000 / 48
>                   7   0/1  111000 / 56
>
>    The above all follow the recommended DSCP values from [RFC2474]
>    except for the placing of priority 1 below priority 0, as specified
>    in [802.1Q], and for the DSCP value of 000010 binary for low effort
>    as recommended in [LEphb].
>
>
> I don't really understand this change. Now you have priority 1 which
> points to the proposed new mapping for Lower than Best Effort, being the
> least prioritized. Then priority 0 intended to be more prioritized is CS1
> which may still be lower than best effort (thus basically same as prio 1),
> or end up being receiving a more prioritized PHB than best effort. Why
> isn't 0 mapped to best effort (000000)? That appears that it would provide
> a more consistent behaviour.
>

I think that in updating this section there was too much focus on updating
the DSCP value for low effort and not as much focus on the rest. We'll see
what we can do to straighten this out.

> MTU and Fragmentation
> ---------------------
>
>
> With the changes introduced and the no middlebox applicability it looks
> like the issues are resolved.
>

Thanks

> Zero Checksum:
> --------------
>
>
> The IPv6 zero checksum section (5.4.2) looks good.
>

Thanks

> TCP Encapsulation issue
> -----------------------
>
> So the TCP encapsulation section (5.6) is significantly improved, but some
> comments:
>
>
> "All packets in a particular TCP stream SHOULD use the same DSCP
> codepoint or packet re-ordering may occur."
>
>
>
>
>
>
> I think that SHOULD is a MUST. You get no benefit from trying to send
> different packets in a TCP connection with different DSCPs. It will really
> only result in reordering, and additional retransmissions. Also the TCP
> stacks Head of line blocking will also not really allow any received
> payload to be released early. Thus, only downsides and now gain to use
> different DSCPs for a single connection.
>
> "It is RECOMMEDED that
>    at least two TCP connections be provided, for example one for
>    priority 6 and 7 traffic and one for priority 0 through 5 taffice, in
>    order to insure that urgent control traffic, including control
>    traffic related to establishing and maintaining adjacency, is not
>    significantly delayed by lower priority traffic."
>
> I think this is fine, there is one potential issue here, but likely not a
> significant. If the high priority traffic is very sparse, then the TCP
> connection marked with a higher priority DSCP may have a very small
> congestion window. So when the high priority traffic is to be sent it will
> in fact be delayed more by the TCP stacks congestion control than the lower
> priority that has a larger window established. However, that is only really
> an issue if the high priority traffic is multiple packets that comes in
> burst. Also, if there is on path contention the lower priority traffic may
> still not make it even if sent earlier, so not using a high priority DSCP
> is not really an option either.
>
> I think the IP/TCP/Trill and TCP header figures are simply non-relevant
> here. I think they can give the impression that a receiving node can
> process each TCP packet individually rather than running a full TCP
> implementation here. What is relevant is the framing format that delimit
> each Trill packet sent by TCP.
>

OK.

+--------+-------- // ----+ | Length | TRILL packet   |
>       +--------+----- // -------+
>
> The above figure is a bit confusing, I assume the intention is to indicate
> previous Trill packet with framing and the following one. I think you be
> more explicit about that as an explanation.
>

That figure is intended to be the more detailed expansion of the "Framed
TRILL Payload material" that appears in the IP/TCP/Trill and TCP header
figures shortly above it. More explanation can be added here.


>          If the
>          initial 2 bytes of the TRILL packet are not one of these
>          Ethertypes, then the receiver assumes that framing
>          synchronization has been lost and MUST close that TCP
>          connection.
>
> Yes, this is very important text. I would recommend that it is moved into
> its own paragraph rather than hidden in a figure explanation.
>

OK.


> I am also missing the specification what to do when one has closed the
> connection. I assume that it is to have either part re-initiate. But, is it
> the closing party (receiver) that should do this, or the sending part? As
> TCP connections are bi-directional, but I assume that these are used only
> unidirectionally, and thus it matters who should establish them in relation
> to signalling? Congestion Control
>

There should probably be a paragraph on TCP connection establishment
including how to handle this case.

> Flow and ECMP
> -------------
>
> Issues resolved in the more limited applicability scope.
>

Thanks.

> NAT and TRILL over IP:
>
>
> I think it is fine with the WG saying that NATs are out of scope for
> Trill. However, I really think you need to say this upfront in the document
> that this specification relies on that there are no NAT in the path between
> the trill nodes using this specification. And be explicit if there would be
> NATs then things will not work as intended.
>

OK

Thanks,
Donald
===============================
 Donald E. Eastlake 3rd   [phone] (cell)
 155 Beaver Street, Milford, MA 01757 USA
 [e-mail]

Cheers
>
> Magnus Westerlund
>
> ----------------------------------------------------------------------
> Media Technologies, Ericsson Research
> ----------------------------------------------------------------------
> Ericsson AB                 | Phone  [phone] <+46%2010%20714%2082%2087>Torshamnsgatan 23 <https://maps.google.com/?q=Torshamnsgatan+23&entry=gmail&source=g>           | Mobile [phone] <+46%2073%20094%2090%2079>
> SE-164 80 Stockholm, Sweden | mailto: [e-mail]
> ----------------------------------------------------------------------
>
>