Re: Interface reality check
Thanks for chiming in! I have great respect for your opinion. You have a few
questions directed at me in your comments. I'll address those below.
Steve Haddock wrote:
> Hi Rich, Hi Ben!
> New blood? Why would you want new blood? You're doing so well! Or do you
> really just want fresh meat?
> OK, I'll chime in. The pictures and the discussion have been great. They
> answered a couple of questions that were nagging me, although these
> questions never came up directly in any reflector message that I saw. Just
> to make sure I'm drawing the right conclusions, these were my questions:
> 1) Is it necessary for a PCS/PMD to carry /A/K/R/? I believe the answer is
> no. When this whole issue came up I shared Ben's assumptions that /A/K/R/
> would not be carried across a 64/66 link, but Rich's passion for carrying
> them end-to-end made me wonder whether it was necessary for correct
> operation of XAUI in some case that I was missing. The pictures have
> convinced me that there is no absolute requirement for carrying /A/K/R/
> end-to-end. I believe this is true even if the link is a 8b10b encoded WWDM
> link that uses /A/K/R/ in a similar manner to XAUI. Things will operate
> properly if /A/K/R/ is translated to /I/ and then regenerated as an
> independent /A/K/R/ pattern at each XAUI-WWDM-XAUI interface (although I
> doubt any implementations would actually do it this way).
I agree with this since it's reflected in the pictures (e.g. page 2). Page 2
shows that PCS/PMA transport of /A/K/R/ is not required. Page 2 shows that
PCS/PMA transport of /S/d/I/E/RF/BL/ and possibly /O/ is required.
> 2) Is it beneficial for a PCS/PMD to carry /A/K/R/? Rich's point of view
> is that it is simpler if /A/K/R/ does not have to be translated to /I/ at a
> XGXS-to-PCS boundary, but I would argue that simplicity is not the same as
> beneficial. Since a PCS-to-XGXS boundary cannot assume there will be
> /A/K/R/ coming across the link (there may not be a XAUI at the other end),
> it has to generate /A/K/R/ patterns itself. At that point there is no
> benefit to having /A/K/R/ carried across the link.
> 3) Is there a disadvantage to carrying /A/K/R/ across a PCS/PMD? The only
> disadvantage that I see is that it consumes contol-code space in the PCS
> encoding scheme. If the control-code space is available, this is a very
> minor disadvantage.
Good point. /A/K/R/ is specifically defined as a column of /A/s or /K/s or /R/s.
This uses up 3 codes. However, these codes or their corresponding translation
should not be used by any other PCS encoding scheme. In addition, an appropriate
hamming distance between these codes and other defined codes should be observed.
I agree that the disadvantage is minor and probably negligible.
> So in the end I'm ambivalent about whether or not /A/K/R/ get carried across
> a PCS/PMD or not, but I do feel rather strongly about how this translates
> into requirements in the specification. What I would propose is:
> a) A PCS/PMD is not required to propagate /A/K/R/ that it receives when
> (and if) attached to XAUI.
> b) The RS receiver and any PCS receivers shall treat any
> unknown/unrecognized control symbol as an idle when received during the
> Inter-Packet-Gap. Unknown control symbols received in a frame should be
> treated as an error, but between a valid end delimiter and a valid start
> delimiter there is no particular need or benefit to such error checking. An
> over-constrained specification is bad for interoperability and restricts
> degrees of freedom that may become important someday. By the way, encoding
> schemes that can distinguish between truly invalid symbols and
> unknown-but-still-valid control symbols should count invalid symbols as
> errors. This allows detection of poor signal quality on an idle link.
Agreed. Cool! 8B/10B plug: 8B/10B supports the detection of poor signal quality
on a link regardless of the information being transported. In fact simple
receiver circuitry can determine the BER at the decoder in real time. Try that
with a scrambled code ;_)
> I think it is cleaner if /A/K/R/ is not allowed outside of XGXS boundaries,
> but since I think an RS receiver should be tolerant of any control symbols
> received in an IPG I'd be willing to compromise on allowing an XGXS to
> propagate /A/K/R/ rather than having to translate them to /I/.
> I think there is a bigger issue that came out in Rich's pictures: /O/. As I
> said above I'm in favor of permissive receivers, so I don't care if a /O/
> shows up in an IPG and gets treated as an /I/. It is a very different thing
> to say that /O/ must be carried end-to-end across any PCS/PMD, which is
> implied in Rich's pictures. This would mean that every PCS/PMD would have
> to know how to encode /O/, which then means we have to know how many flavors
> of /O/ there are, etc. Who is going to define this? I hope what Rich
> really means is that there may common components between 802.3ae and Fibre
> Channel so there may be parts that know how to translate ordered sets from
> XAUI to 64/66, but there is no requirement for every 802.3ae PCS to know how
> to do this and it is OK for any 802.3ae device that doesn't understand Fibre
> Channel ordered sets to treat them as idles.
Your assumption of what I mean is 100% correct. FC Ordered Sets would be
recognizable by an FC-1 element. The same codes received by an Ethernet RS would
be treated as Idles. However, they wouldn't be generated by an Ethernet MAC
transmitter, RS, XGXS or PCS in the first place. Common components would
transparently transport both.
> -----Original Message-----
> From: Rich Taborek [mailto:rtaborek@xxxxxxxxxxxxx]
> Sent: Saturday, April 01, 2000 12:02 PM
> To: HSSG
> Subject: Re: Interface reality check
> We're getting much closer. I have comments in the following areas:
> Code Space
> We seem to agree on reserving some code space for future use.
> Supporting Fibre Channel in IEEE 802.3ae is not relevant. However, it is
> impossible the overwhelmingly strong market acceptance of gigabit
> interconnects, the exponentially growing demand for data transport,
> and the magical convergence speed of ~10 Gbps for the LAN, MAN/WAN and
> IEEE 802.3ae has decided that addressing MAN/WAN applications is in its best
> interest. WAN support is written into our PAR and included in our
> objectives. It is clear that supporting the MAN/WAN means supporting SONET
> directly to some, simply providing Ethernet access to the MAN/WAN to others,
> and building out all Ethernet MAN/WANs to still others. I believe that
> either the IEEE 802.3ae or vendors are just going to bet on one of the
> three MAN/WAN support scenarios. My own analysis of the strategies and
> product lines of large companies such as Cisco and Nortel as well as VC
> investments in the past few years clearly show the billions of dollars are
> being poured into all three. Getting back to code space, I believe that
> 10 GbE needs to be architected in a flexible enough manner to support
> MAN/WAN directions.
> The Storage Area Network business and applications have been hard to ignore
> in 1999 and momentum is still growing. Fibre Channel has already endorsed a
> roadmap which includes 10 Gbps and 40 Gbps interfaces for the SAN on a
> timeline in sync with MAN/WAN timelines. Many SAN applications also require
> MAN/WAN connectivity for applications such as remote backup, disaster
> recovery, and storage sharing between remote corporate sites. Some Ethernet
> vendors have also been eyeing the SAN space as a huge and easy business
> opportunity to put under their wing. Once again, the direction is clear:
> SANs are here to stay. Getting back to code space again, I believe that
> 10 GbE needs to be architected in a flexible enough manner to support SAN
> All these 10 Gbps LAN, MAN/WAN and SAN links need to be sourced somewhere.
> Enter the server, whether it be a low-cost high volume or mainframe
> server. One or more 10 Gbps and higher connections will be directly
> attached to these servers in the near future, requiring internal 10 Gbps
> and higher internal I/O busses. This is where interfaces like
> InfiniBand(TM) and PCI-XXX enter the picture (yeah I made PCI-XXX up :-).
> It just so happens that InfiniBand and 10 GFC and 10 GbE requirements for
> code space are pretty similar when you get right down to it.
> "Brown, Ben [BAY:NHBED:DS48]" wrote:
> > Rich,
> > Rich Taborek wrote:
> > >
> > > > Page 1: I agree with everything here except perhaps the /O/.
> > > > I'd like to understand more about why this is needed.
> > >
> > > I've included the /O/, representing "other" in support of other
> > > standards such as Fibre Channel and InfiniBand with common parts,
> > > for other OAM&P functions for Ethernet WAN access and WAN
> > > applications, and for other unforeseen extensions to 10 GbE
> > > since I believe that we shouldn't assume that we have all link
> > > control functions covered. Both 8B/10B and 64B/66B proposals
> > > currently support additional control codes which may be required
> > > to provide /O/ support for the purposes described above.
> > I agree the code space is available and that it should be held
> > reserved. I think a healthy debate is yet to be waged on the virtues
> > or follies of explicitly supporting other standards (Fibre Channel
> > and Infiniband) within "ae".
> > > > Page 2: I agree with everything here except the following:
> > > > This applies for any serial PCS, 64b/66b or other. I think
> > > > this is the picture that I envisioned when I started this
> > > > thread and wat I understand is the picture that Mr. Rick
> > > > Walker is in support of based on his most recent comments.
> > > > The WWDM encoding may be identical to XAUI or may be
> > > > something else completely. Either way, I'd like to
> > > > see this specified with the encodings supported/
> > > > required by RS only. Even XAUI is specified this way.
> > > > XAUI requires /A/K/R/ for a further encoding of the
> > > > IDLE stream but it takes /I/ as input (along with all
> > > > the others /S/T/d/E/RF/BL/). Our debate is where the
> > > > /A/K/R/ gets stripped off after XAUI.
> > >
> > > I'm not sure I understand exactly what you're disagreeing with on page 2
> > > since its intention was to show your view: "Serial PHY, 64B/66B PCS,
> > > XGXS never forwards /A/K/R/." Page 2 clearly show /A/K/R/ being stripped
> > > off and translated back to /I/ by the XGXS adjacent to the PCS in Device A.
> > >
> > > Our debate has also focused on only the Serial PHY and only on 64B/66B
> > > coding. I'd like to keep our debate focused on these elements only.
> > > I believe that this coding debate is applicable to and independent of
> > > other PHY's and encodings, but I'd like to focus on the PHY/PMD we
> > > started with, especially since it is endorsed by a significant cross
> > > section of IEEE 802.3ae Task Force members.
> > >
> > > Can you please explain your disagreement with page 2 with respect only
> > > to the Serial PHY with a 64B/66B PCS.
> > With respect to a Serial PHY using 64b/66b, I am in full agreement
> > with page 2. I was merely pointing out that this can also apply to
> > other serial PCS proposals in addition to 64b/66b and that WWDM is
> > a bit off the beaten path. These discussions probably deserve their
> > own thread.
> > > > Page 3: I think that this is the picture you've been in
> > > > support of and the one that the current 64b/66b proposal
> > > > describes. I don't agree with this picture.
> > >
> > > If you'll allow me to twist your words: I believe that you agree that
> > > the picture accurately represents my previous view of XAUI/XGXS and
> > > PCS operation for the Serial LAN PHY with 64B/66B encoding. Your
> > > disagreement is with this mode of operation, not the picture. I know
> > > it's a minor point, but I am trying to accurately portray our debate
> > > in pictures since words are apparently not working.
> > Again, you are absolutely correct. I disagree with this mode
> > of operation. Because this picture represents this mode of
> > operation, I don't particularly like it but my distaste is
> > based solely on the mode of operation that it represents, not
> > the picture itself. I think it is a great picture to describe
> > your previous view of XAUI/XGXS and PCS operation.
> Great! that was my intention, to picture what you disagreed with and allow
> others to see the same.
> > > > Page 4: A minor twist to page 3.
> > >
> > > Wow! This comment boggles my mind!
> > >
> > > This picture is the heart of the presentation and illustrates a solution
> > > to problems exemplified in page 2, call it Ben's picture, and page 3,
> > > call that one Rich's picture. I must not have done a good job on the
> > > picture in page 4 or maybe they all look too similar.
> > >
> > > This page shows control codes being generated by the transmitting RS and
> > > modified by the transmitting XGXS, if present. Control codes are then
> > > transparently transported by the PCS and over the medium. All control
> > > codes not specified by 10 GbE are translated to Idles by the RS.
> > > The latter translation occurs in a manner analogous to that of the
> > > 1000BASE-X PCS Receiver.
> > Again, the picture is a fine one that, I believe, accurately
> > represents your current view of XAUI/XGXS and PCS operation.
> > I guess I should have been more explicit. I'm not against the
> > incremental step that this page shows. This incremental step
> > of having the RS treat everything it doesn't recognize as an
> > /I/ is fine. I merely disagree with the base content that was
> > carried forth from page 3.
> Now we're getting to the root of the problem. We already agree that multiple
> control codes must be transported by 64B/66B and the medium. These include
> /S/d/I/E/RF/BL/ and we even agree that there are some good reasons to
> transport /O/. I'm looking for the simplest way to support all required and
> optional 10 GbE sublayers and interfaces in a PMD independent fashion.
> I would do the same for an SLP-based PCS for the Serial LAN or WAN PHY, for
> WWDM based on 8B/10B and for WWDM based on an alternate PCS. Having the RS
> Receiver simply treat all undefined codes as Idles the simplest scheme I
> could come up with. This scheme is capable of supporting all proposed PCS
> and interface codes assembled in any combination for the 10 GbE link.
> I believe that you'll find that the "base content" that you disagree with
> may be different for different PCS and optional interface codes. The scheme
> outlined on Page 4 is "base content" independent.
> > > > Page 5 & 6: I don't see any difference between these 2 pages.
> > > > I agree with this.
> > >
> > > I included these for completeness. They illustrate the "reverse" or
> > > Device B to Device A path analogy to Pages 2 and 3, respectively.
> > >
> > > > Page 7: A minor twist to 6 & 7. Though subtle, I don't like
> > > > this but I could probably be convinced by using the same
> > > > arguement that allows the 1000Base-X receive state machine
> > > > to treat "everything else" as /I/.
> > >
> > > I'm still surprised by your comment on page 4 since Page 7 illustrate
> > > the "reverse" or Device B to Device A path analogy to Page 4. Your
> > > statement that "to treat "everything else" as /I/" is exactly what
> > > I had in mind. I'm offering this as the solution to our debate. I'm
> > > open to other solutions which can resolve our debate in a simple
> > > manner.
> > As I tried to clean up earlier, I agree that this incremental
> > step is a good one (having the RS treat everything it doesn't
> > recognize as an /I/). I simply disagree with carrying /A/K/R/
> > outside the boundary of the XGXS blocks.
> > The bottom line is that we're still in disagreement with the
> > fundamental aspect of this proposal. The pictures are great,
> > the addition to the RS is great. I simply don't think /A/K/R/
> > should be carried outside the boundaries of the XGXS blocks
> > and you think that it's okay if it does. I don't seem to have
> > changed your mind and you don't seem to have changed mine.
> > I think we could have gotten to this point about a month
> > faster if we were drawing on napkins, sitting in nice comfy
> > chairs drinking Guiness.
> No doubt. Thanks for boiling down your real concern though. I believe the
> pictures have helped.
> Transporting /A/K/R/ through 64B/66B, SLP, SUPI or any other Serial or WWDM
> PCS code, given the other mandatory requirements to transport control codes
> is supported in current proposals (e.g. 64B/66B). It is simple to implement
> (I can prove that). I'd like to understand your reasons for not carrying
> /A/K/R/ outside the boundary of XGXS blocks? Do you object specifically to
> the actual 36-bit encodings of the /A/, /K/ or /R/ columns? Why should
> these be treated any differently than a 36-bit encoding for /I/?
> Personally, I think we're down to nit-picking. I've offered up a solution to
> get away from the nit picking. I'd really like to hear from you and others
> from an implementation perspective regarding this solution.
> > > > I don't think any of my responses have surprised you. At this
> > > > point in the thread, I think we know where each other stands.
> > > > It would be interesting to hear a few others weigh in on this.
> > > >
> > > > Thanks for the nice pictures. This should make involvement by
> > > > the rest of the group much easier.
> > >
> > > That was the intention. Anyone else out there want to offer their
> > >opinions?
> > Please, let's get some new blood in on this discussion! I feel
> > like we're in a vacuum.
> I agree. We either need more participation, or more Guinness :-)
> > Ben
> > --
> > -----------------------------------------
> > Benjamin Brown
> > Router Products Division
> > Nortel Networks
> > 1 Bedford Farms,
> > Kilton Road
> > Bedford, NH 03110
> > 603-629-3027 - Work
> > 603-624-4382 - Fax
> > 603-798-4115 - Home
> > bebrown@xxxxxxxxxxxxxxxxxx
> > -----------------------------------------
Richard Taborek Sr. Phone: 408-845-6102
Chief Technology Officer Cell: 408-832-3957
nSerial Corporation Fax: 408-845-6114
2500-5 Augustine Dr. mailto:rtaborek@xxxxxxxxxxx
Santa Clara, CA 95054 http://www.nSerial.com