[802.3ae_Serial] Updated link models
Colleagues,
At last I have carried out the actions put on me a long while ago to update
the link model.
Please find two spreadsheets.
At
http://www.ieee802.org/3/ae/public/adhoc/serial_pmd/documents/10GEPbud2_4_2D
3.1v3.xls
there is a copy of our "official" link model 2.4.1 with minimal changes to
align it with IEEE 802.3ae draft 3.1. The changes are listed in green, on
the Notes page. The major ones are:
* Deterministic jitter replaces receiver eye penalty
* Reflection noise is included
At
http://www.ieee802.org/3/ae/public/adhoc/serial_pmd/documents/10GEPBud3_1_11
.xls
there is an upgraded link model.
The major changes over 2.4.1/2.4.2 are:
* Improved layout
* OMA-oriented
* "Box level" or system level spec concentrates on eye margin at TP2,
and penalties at the eye centre at TP3. TP4 and "eye corner" penalties are
"for information only".
* Complete re-write of stressed eye sensitivity as discussed on
conference calls. The idea is that the stressed and nominal sensitivity
tests would represent equal levels of performance as each other, with equal
margins. Note that now, these two quantities track each other and if the
margin is significantly wrong, both will be wrong together (also note that
at present the nominal sensitivity is miscalculated: this does not affect
anything else in the model).
* Deterministic jitter integrated into model
* Reflection noise calculated in model. There is a rather arbitrary
"Reflection Noise factor = 0.6".
* Calculates stressed eye vertical eye closure penalty automatically
for copying into standard.
* Revision to RIN penalty following Dubravko Babic: The effective
Signal to Noise ratio (which drives the bit-error rate at TP3/4) caused by
RIN is ISI-reduced eye opening / bandwidth-filtered noise. The ISI part had
always been included as its own penalty and had been used as the numerator
in the signal-to-noise calculation for baseline wander (BLW), but had been
forgotten as the numerator in the signal-to-noise calculation for RIN. In
this model the RIN variance still goes down with length or reduced bandwidth
(ISI) as before. But ISI deteriorates faster, going to infinity
(deterministically closed eye) at some system bandwidth. Therefore, like
baseline wander, the penalty goes up as ISI increases, going to infinity
(error floor) when the effective Signal to Noise ratio falls below 7. None
of this has anything to do with attenuation, which cancels out. The "cross
penalty" as before accounts for the superlinear combination of penalties.
See "notes" page for full details, and see "Notes for use" near the bottom
of that page for use of eye height to set eye margin and (for 1550 nm only)
vertical eye closure penalty (VECP) which I think is equivalent to
Transmitter and dispersion penalty (TDP).
*** It is important to note that the numbers are not set to the "worst case"
but to values we are familiar with. In some cases the eye margin is
negative (eye height <50%, worse than worst), in other cases it is positive.
Worst case may be with very little RIN.
Bugs and errors aside, I am recommending this new model for adoption by 10G
Ethernet (the parameters like reach, jitter, Tx power, sensitivity which
populate the spreadsheet are separate issues).
Here's one bug: the nominal receiver sensitivity is miscalculated, because
of an imperfectly implemented change in basis wavelength e.g. 1290 to 1265
nm. The values in the draft standard are probably more consistent. This
error does not affect anything else in the model.
The following notes are a discussion of what I think the model tells us
about the draft standard. It is important to realise that changing the
model does not make the optics any less satisfactory. What it does do is
spell out more clearly what the causes of performance are, rather than
having some effects unquantified or lumped together in "margin". The
experimental work of the PMD feasibility ad hocs will help resolve any
issues.
The 850 nm cases are affected by the more pessimistic RIN calculations.
This is not quite fixed by tightening the RIN spec; see "Notes for use" to
understand why tightening the RIN spec unduly is pointless. Slight changes
in reach, fibre bandwidth or sensitivity (budget) would in most cases make
the margin positive again.
850 nm, new MMF also suffers from mode partition noise with as much as 0.4
nm RMS spectral "width". As can be seen by inspection of Table 52-8-10,
GBASE-S triple trade off, penalties increase rapidly with high spectral
width. I expect with the new model, the curve will be steeper.
LX4 on SMF could suffer from reflection noise, because an extinction ratio
as low as 1 dB is allowed per draft 3.1.
1310 nm serial shows a slightly negative margin. In SSIG, a long time ago,
this PMD was built from the receiver upwards. We need to revisit that
process: set the receiver sensitivity to a reasonable point, overcome the
losses and impairments, write down the necessary Tx power. It looks like we
may have to go back to the 10 dB link budget we started with, if we don't
change the attenuations.
The 1550 serial case seems to have about 1 dB of margin (you need to adjust
parameters to set the eye margin and VECP to see this).
A discussion of what margin is appropriate would be worthwhile, when we can
share some measured performance reports. One could argue that this model
codifies some things which were previously covered by "margin", and no more
is necessary.
Best regards,
Piers Dawe