Re: [HSSG] Higher speed trade offs
Hi All
Chris, thanks for the numbers on skew, they serve as a useful starting
point. I also reviewed the presentations at this week's meetings.
Unfortunately I have teaching commitments and cannot attend.
I did want to point out some results based on some initial analysis and
some review of the presentation material.
It appears to me that any link aggregation technique will suffer from
the BER of the underlying PHY technology. For example a 100G solution
constructed from 10GBASE-T will exhibit a BER of 1e-12 unless we add
additional complexity. This was the focus of a thread on BER targets a
couple of weeks ago.
One interesting thing to note is that if we do aggregation above the
MAC, adding any additional FEC may be difficult. As I understand it,
typically the MAC will not output Ethernet frames with invalid CRCs?
Also at this point all soft-information has been lost so only
hard-decoding would be possible.
In several presentations people suggest an aggregation scheme
implemented above the PHY. In that case we could consider using an FEC.
If you presume we can implement a high-rate code capable of correcting
any single bit error in one of the data fragments then the BER changes
from 1e-12 to
1 - (1-1e-12)^N - N*(1-1e-12)^(N-1)*1e-12
So for N=1024 (using the 1Kb skew length Chris suggested) we get a new
BER of about 1e-19. Of course we have to assume bit errors are IID and
this may not be the case if there is already an underlying FEC.
However the interesting initial conclusion is that if we do decide to do
link aggregation it should be possible to utilize a high rate code that
helps us achieve a reduced BER target.
It would be useful to develop some error models for some of the proposed
underlying PHY schemes. Does anyone have any such models and be willing
to share them with the group?
thanks
Stephen
------------------------------------------------------------------------
Dr. Stephen Bates PhD PEng SMIEEE
Department of Electrical and Computer Engineering Phone: +1 780 492 2691
The University of Alberta Fax: +1 780 492 1811
Edmonton
Canada, T6G 2V4 stephen.bates@xxxxxxxxxxxxxxx
www.ece.ualberta.ca/~sbates
------------------------------------------------------------------------
Chris Cole wrote:
> Stephen,
>
> The main stream SMF application at 10G is 10km (10GBASE-LR,) which uses
> un-cooled optics.
>
> Assuming that 100Gb will use un-cooled optics for low cost, this
> requires CWDM. A 100nm window is one alternative that can be used for
> skew calculations. The dispersion across the window is approximately
> 10ps/nm/km, so we get 10nsec after 10km. That's 1000 bits, which will
> get split up among however many CWDM channels are proposed.
>
> An 80km application will require the use of cooled optics. A 10 channel
> 200GHz DWDM grid will be 16nm, so that also gives about 10nsec after
> 80km.
>
> So in round numbers, you can use 1Kb as a starting point.
>
> Chris
>
> -----Original Message-----
> From: Stephen Bates [mailto:stephen.bates@xxxxxxxxxxxxxxx]
> Sent: Thursday, September 14, 2006 10:58 AM
> To: STDS-802-3-HSSG@xxxxxxxxxxxxxxxxx
> Subject: Re: [HSSG] Higher speed trade offs
>
> Hugh and others
>
> We have been thinking about 100G for a while and there does seem to be
> potential for maximum-likelihood detection. This is somewhat similar to
> some of the PRML work done in hard-drives, which also tend to use binary
>
> signaling.
>
> Also, assuming a binary signaling scheme, the constellation size is not
> too much larger than the 625 used in 1000BASE-T (trellis code with 5.5dB
>
> coding gain) and a lot smaller than the 65536 used in 10GBASE-T (LDPC
> with about 10-12db coding gain depending on your decoder). The latency
> through the 1000BASE-T decoder was only about 20 symbol periods so its
> impact on latency at 10G would be small compared to the potential bulk
> skew delay (I presume).
>
> If there is interest from the group for an FEC we would like to develop
> these ideas further. Also, can anyone tell me what the worst case bulk
> skew is likely to be in some of these multi-wavelength schemes (as a
> function of cable length)?
>
> cheers
>
> Stephen
>
> ------------------------------------------------------------------------
>
> Dr. Stephen Bates PhD PEng SMIEEE
>
> High Capacity Digital Communications Laboratory
> Department of Electrical and Computer Engineering Phone: +1 780 492 2691
> The University of Alberta Fax: +1 780 492 1811
> Edmonton
> Canada, T6G 2V4 stephen.bates@xxxxxxxxxxxxxxx
>
> www.ece.ualberta.ca/~sbates
> ------------------------------------------------------------------------
>
> Hugh Barrass wrote:
>> Stephen,
>>
>> Regarding the FEC & latency - a FEC that is designed to exploit the
>> transverse dimension (i.e. correlation between channels) would not
> need
>> to add significant latency. The FEC block size (or equivalent) need
> only
>> be the same order as the maximum skew between the channels. This will
>> govern the minimum latency for a non-FEC channel in any case. At its
>> simplest, a Trellis code could be applied across the channels with an
>> additional latency of ~ 1 code block. A cleverly designed maximum
>> likelihood code (is anybody in Alberta or Cork working on that? :-)
>> could offer similar gain with lower overhead. In particular, the
> optical
>> channel with binary signaling offers a much smaller problem matrix
> than
>> 1000BASE-T multi-level FEXT/NEXT channels.
>>
>> The processing overhead is largely dependent on the amount of state
> and
>> to a first approximation would scale with latency.
>>
>> Hugh.
>>
>> Stephen Bates wrote:
>>
>>> Hi Hugh and others
>>>
>>> I have been following this mailing list with interest and wanted to
>>> comment on Hugh's statement in his email.
>>>
>>> "It strikes me that if the sources and destinations of many carriers
>>> are co-located and correlated then coding can eliminate inter signal
>>> interference."
>>>
>>> I am trying to understand what the advantage is of merging 10 10G
>>> channels into one 100G channel versus keeping the 10G channels
>>> separate. It seems to me all the buffering and SAR requirements
>>> required are only of value if we do take advantage of all the
>>> dimensions. Obviously this is something we've done in 1000BASE-T and
>>> 10GBASE-T by running some kind of FEC over all four dimensions.
>>>
>>> However we've already had a discussion on how FEC adds latency and
>>> that may not be acceptable in short-haul applications. Also, decoding
>
>>> a 10 dimensional code would not be trivial, though the potential
>>> coding gain would be large, allowing dense packing of wavelengths.
>>> Also, if there is significant correlation across the
>>> dimensions/wavelengths we can take advantage of that using
>>> maximum-likelihood detection techniques. Again the complexity and
>>> latency become issues. However the maximum likelihood approach is
>>> interesting in that it can be utilized without compensating for any
>>> bulk skew mis-match between the dimensions/wavelengths.
>>>
>>> I look forward to seeing how this work develops.
>>>
>>> Cheers
>>>
>>> Stephen
>>>
>>>
> ------------------------------------------------------------------------
>>> Dr. Stephen Bates PhD PEng SMIEEE
>>>
>>> High Capacity Digital Communications Laboratory
>>> Department of Electrical and Computer Engineering Phone: +1 780 492
> 2691
>>> The University of Alberta Fax: +1 780 492
> 1811
>>> Edmonton
>>> Canada, T6G 2V4
> stephen.bates@xxxxxxxxxxxxxxx
> www.ece.ualberta.ca/~sbates
> ------------------------------------------------------------------------
>>> Hugh Barrass wrote:
>>>
>>>> Andrew and others,
>>>>
>>>> It often amuses me that technical principles from one field of
>>>> invention seem to leak into other fields. The mechanism that you
>>>> suggest strikes me as very similar to Discrete Multi-Tone
> modulation,
>>>> used in DSL. There are some considerable advantages of compact multi
>
>>>> carrier systems over higher baud rate single carrier systems. I
> guess
>>>> it's only a matter of time before someone comes in (or back) with
>>>> optical multi-level signaling to make the matrix complete :-)
>>>>
>>>> Not being an optical expert allows me the freedom to look at this
>>>> from the outside and to suggest some ideas that may (or may not) be
>>>> completely hopeless. Has anyone considered the use of FEC codes
>>>> designed to correct errors caused by ultra-fine WDM spacing? It
>>>> strikes me that if the sources and destinations of many carriers are
>
>>>> co-located and correlated then coding can eliminate inter signal
>>>> interference.
>>>>
>>>> Isn't communications theory fun? :-)
>>>>
>>>> Hugh.
>>>>