Thread Links Date Links
Thread Prev Thread Next Thread Index Date Prev Date Next Date Index

Re: [HSSG] Trowbridge_01_0907.pdf comment about 49.2.4.3



Andre,
If backplane were the only 40 GbE interface we were planning to build, this might be an easy choice.
 
But as it is, there is also merit in common architecture across the PMDs. And there is this sticky question about what constitutes a transparent service over OTN to worry about, and how one might deal with different Ethernet PMDs at the ingress and egress if there are significant architectural differences across PMDs.
Regards,
Steve


From: Andre Szczepanek [mailto:a-szczepanek@ti.com]
Sent: Thursday, September 27, 2007 7:01 AM
To: Trowbridge, Stephen J (Steve)
Cc: STDS-802-3-HSSG@LISTSERV.IEEE.ORG
Subject: Re: [HSSG] Trowbridge_01_0907.pdf comment about 49.2.4.3

Steve,

Sorry, I made the mistake of replying to a message before emptying my mailbox !

I think a simple round-robin distribution would work well, as long as lanes are uniquely connected.
This is not without precedent - XAUI has no lane identifiers and relies on correct lane connection.
There is merit in a simple scheme that does not require alignment marker insertion.

Regards Andre


Trowbridge, Stephen J (Steve) wrote:
Andre,
I think I already mentioned that this possibility occurred to me in a
previous reply I wrote to Pat. Of course one would need to take care of
how the sequence of 64B/66B blocks get mapped to FEC codewords. If we
sequentially number the 64B/66B blocks 1, 2, 3, 4, 5 ..., then block 1
is the 1st block in the FEC codeword for lane 1, block 2 is the 1st
block in the FEC codeword for lane 2, block 3 is the 1st block in the
FEC codeword for lane 3, block 4 is the 1st block in the FEC codeword
for lane 4, block 5 is the 2nd block in the FEC codeword for lane 1, and
so on for the 128 64B/66B blocks that fit into the four corresponding
FEC codewords sent on each lane.
Regards,
Steve

-----Original Message-----
From: Andre Szczepanek [mailto:a-szczepanek@TI.COM] 
Sent: Thursday, September 27, 2007 1:53 AM
To: STDS-802-3-HSSG@LISTSERV.IEEE.ORG
Subject: Re: [HSSG] Trowbridge_01_0907.pdf comment about 49.2.4.3

Steve,

Your statement : "No such (alignment) marker appears in any 10G serial
PHY" is not strictly true.

The 10GBASE-R PCS FEC layer standardized by Backplane Ethernet imposes a
2112 bit FEC block structure onto the serial bit stream. FEC block
boundaries can be used as alignment markers to allow ~ +/- 1000bd of
interlane deskewcapability.

The downside of FEC is of course the additional error correction latency
(2112+ bd), however if FEC framing was supported without correction then
this could be avoided (Currently the standard requires that if you
support FEC you must error correct).

Regards
Andre Szczepanek (TI)

Trowbridge, Stephen J (Steve) wrote:
  
Hi David,
Understood that the state machine is the (most) normative part of the 
specification, but I don't think you want to rely on things that are 
buried deep in the state machine to provide the appropriate safeguards
    

  
about where one needs to be careful in future evolution of the 802.3 
standard so as not to break a mapping into OTN. No doubt that the 
state machine must be designed so that it interworks properly with the
    

  
mapping into OTN that will be standardized concurrently in ITU-T. But 
making sure that it works with the initial version of the standard and
    

  
making sure that nobody breaks it later both need to be covered.

Regarding the 40G PHY, we haven't (and aren't allowed to yet, 
according to the rules) chosen a solution. I think that it is widely 
assumed that the 40G PHY will be a 4-lane interface with each lane 
bearing a lot of resemblance to a 10G serial PHY, but it WILL be 
different. For one thing, the lane of the 40G PHY needs to have lane 
alignment markers for deskew when combined with the other three lanes.
    

  
No such marker appears in any 10G serial PHY (and the markers for 10G 
parallel PHYs are based on replacing four 10B consecutive codewords 
from an IPG - since I think it is widely assumed that we won't be 
using 8B/10B for the lanes of 40G (which wouldn't correspond to any 
serial PHY for 10G anyway), so the 10G Base-X lane marking strategy is
    

  
unlikely to be used for 40 GbE). Whether the lane alignment markers 
are inserted into a lane by doing a rate adaptation on the MAC 
(stealing enough IPG to make room) and just making the marker look 
like a control block, or whether the lane rate is increased slightly 
to make room for the marker is still TBD. So I agree that it is most 
likely that lanes of the 40 GbE PHY will be quite similar to a 10 GbE 
serial PHY, I do not see any way that the lanes of a 40 GbE PHY could
    
actually BE a 10 GbE serial PHY.
  
Regards,
Steve

-----Original Message-----
From: David Martin [mailto:dwmartin@NORTEL.COM]
Sent: Wednesday, September 26, 2007 1:26 PM
To: STDS-802-3-HSSG@LISTSERV.IEEE.ORG
Subject: Re: [HSSG] Trowbridge_01_0907.pdf comment about 49.2.4.3

Steve,

A couple of comments on this...

Just so you're aware, it's 802.3 practice that a state machine is the 
definitive description of a function and takes precedence over any 
related textual descriptions. Textual descriptions are intentionally 
minimized.

Re your last paragraph, I'm not sure I follow the latter part. I would
    

  
expect 40GigE to use multiple serial 10G PHYs rather than multiple 
multi-lane 10G PHYs.

For example, the 10GBASE-KR and 10GBASE-SR serial PHYs both include 
the
c49 64b/66b PCS, not 8b/10b. And if APL above the PCS was used, then 
those PHYs could be leveraged in their entirety.

...Dave

David W. Martin
Nortel Networks
dwmartin@nortel.com
+1 613 763 3874 (esn 393)
~~~~~~~~~~~~~~~~~~~~~~~~~

-----Original Message-----
From: Trowbridge, Stephen J (Steve)
[mailto:sjtrowbridge@ALCATEL-LUCENT.COM]
Sent: Tuesday, September 25, 2007 6:00 PM
To: STDS-802-3-HSSG@LISTSERV.IEEE.ORG
Subject: Re: [HSSG] Trowbridge_01_0907.pdf comment about 49.2.4.3

Hi Pat,
Yes, I saw this, but I still think some stronger language is warrented
    

  
assuming that we end up relying on a fixed, controlled codeword set as
    

  
a solid "handoff point" between Ethernet and OTN for 40 GbE and ODU3. 
If the mapping from one to the other across standards groups is going 
to depend on it, I don't think you want to trust that everybody goes 
through an extra level of indirection to find every constraint on the 
handoff. It is fairly simple to state the constraints that apply to 
that handoff point clearly in one place, and add the note about the 
relationship to G.709 (assuming that the two standards progress along 
the lines of the transcoding solution to carry one over the other). 
Then you minimize both the chance that someone breaks the transcoding 
with something proprietary, and the chance that the transcoding gets 
broken with the evolution of 802.3 by something inadvertently adding 
something that they didn't know would be a problem because they 
weren't aware of the relationship.

I think the text for 40G PCS will be different from 10G PCS anyway 
because 10G has no multi-lane 64B/66B encoded interface (all of the 
10G multi-lane solutions are 8B/10B encoded). So there is the 
opportunity to make the constraints more explicit for 40G without 
having to tinker with text that applies to previous rates.
Regards,
Steve

-----Original Message-----
From: Pat Thaler [mailto:pthaler@BROADCOM.COM]
Sent: Tuesday, September 25, 2007 2:42 PM
To: STDS-802-3-HSSG@LISTSERV.IEEE.ORG
Subject: [HSSG] Trowbridge_01_0907.pdf comment about 49.2.4.3

I want to respond to the comment made on page 7 of the
trowbridge_01_0907 presentation that says the text about unused type 
field values is too weak.

Steve felt that the language in 49.2.4.3 was too weak to ensure that 
unused values of the type field aren't used - that subclause doesn't 
have a shall statement about transmission and receipt of reserved 
values. The reason a statement doesn't appear there is that the 
behavior is controlled by the state machines and we didn't want to 
double specify it.

Look at the R_BLOCK_TYPE and T_BLOCK_TYPE functions defined in 
49.2.13.2.3. Any block type that doesn't meet the defined types gets 
the value E which causes the transmit and receive state machine to 
turn it into the error block.

Use of reserved type field values is fully covered - It is just done 
in a different location in the Clause.

Pat