Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Matt,
Not to beat on the dead horse, but if you plan to ?line up groups of packets
together and still use long FEC codewords?, this starts to look like a
downstream DBA to me, involving additional packet delay to fit into the next
right FEC word so that it can be sent downstream in the right moment of
time. To me, it seems that this approach would make MAC Control (MPCP in
specific) PHY aware, since right now MPCP will have to know when the right
FEC word starts and ends, when to start downstream transmission and what
specific overhead it will involve. Note also that EPON does not do that
right now, so what you?re suggesting moves the project away from the whole
?EPON Protocol for Coax? approach the project started with into ?let?s do it
the DOCSIS way? approach.
While everything is possible, I question how that would fit into the overall
framework of things in Ethernet. You would not be adding only downstream
scheduler and PHY awareness, but also the need to rapidly change the
effective data rate in PHY (from one FEC word to another) which would need
to be accommodated by in the MAC Control and configured somehow in PHY so
that the idle deletion function removes extra idles (assuming 10G-EPON like
mechanism is employed). There are also all the aspects related with burst
like transmission in downstream to consider, something that is not trivial
to build reliably when we?re talking about synchronizing on a single FEC
codeword (no matter how long it is) and then hunting for a new one to come
at some undefined time in the future. Unless you?re talking about downstream
scheduling, which is a completely new ball game altogether.
In summary, I agree with Ed here. We need much more study in here coming
from folks interested in this solution to evaluate all the merits of such a
proposal and how it would fit into the Ethernet networking model and
specifically, functional separation between MPCP, MAC and PHY. You can do
certain things in DOCSIS, since the concept of MAC and PHY is interleaved
there into a single large box, while in 802.3, layering is strictly
observed. This imposes some constrains on how we can do things here, but
also guarantees scalability and reuse of existing designs. Remember that one
of the reasons listed for EPoC was tapping the Ethernet ecosystem, something
you won?t be able to do if the EPoC design is dramatically different from
anything that has been done for Ethernet to date.
Regards
Marek
From: Matthew Schmitt [mailto:m.schmitt@xxxxxxxxxxxxx]
Sent: Monday, October 08, 2012 23:49
To: STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Ed,
I agree that there will be issues to work out, and we are actively working
through them. However, one point to make clear is that such an approach
does not necessarily require the use of a downstream MAP or Gate message ?
there may be other approaches which mitigate both the latency that results
and the need to add new MPCP messages (and the processing involved in
generating them). Or even if it does involve such a message, it doesn't
necessarily need to be constructed based on packets already received. There
are definitely options and that could address some of these issues.
As for the efficiency issue, the expectation is that by having a very
limited number of profiles, and with the majority of CNUs most likely being
in just a couple of them, you should still generally be able to line up
groups of packets together and still use long FEC codewords. There may be
somewhat of an efficiency hit due to greater use of shortened codewords, but
we believe it can be significantly mitigated.
Many of the other issues you've mentioned are ones we're aware of and
looking into solutions for; again, we'll share those as we get them worked
out. But we believe they're all solvable, and that we should be able to
mitigate the complexity and these issues. As for the issue of how this
interfaces to the Ethernet MAC, that's something I'm less familiar with, and
that we should explore further.
In the end, I agree that we need to make a complete technical proposal that
lays out the answers to these questions, and has backing from a wide array
of people. For now, I just want to get the idea out there, and let folks
know that this is the current consensus among my members.
Thanks.
Matt
From: "Ed Boyd (Edward)" <ed.boyd@xxxxxxxxxxxx>
Reply-To: "Ed (Edward) Boyd" <ed.boyd@xxxxxxxxxxxx>
Date: Monday, October 8, 2012 5:26 PM
To: "STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx" <STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx>
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Hi Jorge,
I agree that more analysis is needed for EPoC. I think that we need to
better understand the penalty and if it is even possible without a radically
new Ethernet standard. As far as DOCSIS and EPoC having the same PHY, it is
a great goal if it can fit with the Ethernet standard and goals. If it has
to diverge here, maybe EPoC is a single downstream profile and DOCSIS
supports multiple profiles. EPoC will have a simpler lower delay downstream
with a fixed data rate pipe that is similar to EPON. DOCSIS could be
configured to run in the simple mode or with multiple profiles. It is too
early to say now but I would imagine that we will run into many examples of
EPoC being a subset of the DOCSIS options.
Take care,
Ed?.
From: Salinger, Jorge [mailto:Jorge_Salinger@xxxxxxxxxxxxxxxxx]
Sent: Monday, October 08, 2012 3:10 PM
To: Ed (Edward) Boyd; stds-802-3-epoc@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Ed,
Not trying to answer all your questions and comments, which I think are very
good, but to share some thoughts?
We went through this same topic for DOCSIS 3.1, and although the conclusions
are likely not all directly applicable in the same way for EPON/EPoC, the
plant aspects should be, even in the case of the path characteristics from
the OCU-CNU vs that of the CMTS-CM. With respect to those conclusions, we
find that the distribution of SNR levels at our entire population of
employed CMs, both within individual nodes and across our entire footprint,
and when evaluating CMs or just MTAs, varies significantly. This would
result on some CNUs being capable of operating at considerably higher
modulation orders while others being limited to lower modulation order. So,
in an effort to maximize overall system performance for all CNUs, we find
that having a small/finite variety of modulation/coding schemes per service
group should help significantly.
Hope that explains the rationale for implementing a small/finite number of
MCS (rather than just one per SG). And, given our goal of making the D3.1
and EPoC PHYs the same, we would like to explore doing the same in EPoC.
There was some preliminary discussion about this during the meeting in
Geneva (see the attached presentation from Andrea), in which data collected
at Comcast was presented and the concept of adaptive modulation was
introduced. We should expand on the analysis moving forward.
Regards,
Jorge
From: Ed Boyd <ed.boyd@xxxxxxxxxxxx>
Reply-To: Ed Boyd <ed.boyd@xxxxxxxxxxxx>
Date: Monday, October 8, 2012 4:19 PM
To: EPoC Task Force <STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx>
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Hi Matt,
I would like to see an analysis on the benefits and the complexity added for
the MCS. There was a presentation with quite a few supporters asking for a
broadcast downstream because it is simple and provides the lowest delay.
From my point of view, I fail to see how multiple profiles will work without
a significant change to the EPON/EPoC system and a huge increase in the
complexity. I see much longer delay and therefore a penalty to the upstream
efficiency (more polling) and CNU cost (more buffering in up and down). I
don?t see how limiting the number of profiles helps much. If you have more
than 1, you need generate a downstream MAP and that is the start of the
trouble. Here are my areas of concern:
1) Downstream Efficiency
To generate the MAP, you need to buffer data and send the MAP in advance.
The MAP must have error correction and it should be in a known position so
you can recover after losing a frame. How do we handle frame boundaries,
MAP location, and symbol boundaries? The long efficient FEC would be
replaced with shorter codes or shortened code words. The FEC would be less
efficient and not as effective, correct? How do we handle variable frame
lengths and the MAP frame location being fixed? It would be inefficient to
stop at frame boundaries?
2) Variable Data Rate to the MAC
The downstream data rate will depend on the destination of the frames, their
mapping within the symbol blocks, and the FEC code word. I think that there
would need to be multiple PHY/MAC interfaces (XGMII) and back pressure
signals but this is a guess. I can?t think of another solution since the MAC
control couldn?t calculate it as we planned for sub-rating.
3) Downstream Delay
A block interleaver must be used since a convolutional would hold packets
between profiles. To do the buffering and sending of the MAP in advance, I
think that we would have 2-to-4 interleaver blocks of delay. I?m almost
sure that it is 4 blocks. If it is a block interleaver versus convolution,
it is twice the delay. If that is the case, the MCS interleaver delay will
be 4 to 8 times more. In my example, we had a 250us interleaver delay so
this change would be 1ms or 2ms delay.
4) Complexity
The interleaver is a huge amount of memory at 5 Gbps but it will be 4 to 8
times larger with the multiple profiles. There is also the complexity of
buffering the packets, calculating the FEC overhead, and then mapping it to
the carriers. To help with efficiency penalty for packets to different
destinations, some have suggested grouping the packets together by profile.
Obviously, this will cause problems with the fixed delay required for the
timestamps unless you add a re-ordering buffer on the receiver. A receiver
will be handling multiple FEC blocks in parallel since it will be receiving
from multiple profiles for multicast and unicast packets.
Thanks,
Ed?
From: Victor Blake [mailto:victorblake@xxxxxxxxxxxxxxx]
Sent: Monday, October 08, 2012 11:27 AM
To: STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Hal,
Is there an assumption being made here that ?impairments? will be
automatically identified and some algorithm will be followed to determine
the mcs(MCSes ?) that will be used ? Might it not be the case that these
would be determined in advance by operator?s administrative settings (aka
configurations).
Just want to ask because it sounds like an assumption to me (because it may
be the case ? I?m begging the question here, that an operator may want to
elect a specific subset of supported MCSes and then presumably have some
level of automation from there.
-Victor
From: Hal Roberts [mailto:Hal.Roberts@xxxxxxxxx]
Sent: Monday, October 08, 2012 1:54 PM
To: STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Matt,
I agree that a number of MCSs should be defined and that a specific set of
MCS attributes could consist of a profile, as you propose. The Channel
Model should inform the choice of MCSs. To do this, the Channel Model
should not be a single set of parameters/impairments but identify a range of
values for each impairment, where necessary.
Hal
From: Matthew Schmitt [mailto:m.schmitt@xxxxxxxxxxxxx]
Sent: Monday, October 08, 2012 12:24 PM
To: STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
As an FYI, if deployed for residential services, especially early on I think
that you might have numbers of devices per OLT port on the order of what
might be done for a DOCSIS service group, achieved by aggregating a number
of OCUs onto a single OLT port. That could easily be on order of 250-500
CNUs per OLT port. Probably less in actuality, but I think we should be
prepared for those types of numbers.
Over time, I agree with Jorge and Ed that it will likely get smaller (by
disaggregating those OCUs onto different OLT ports), probably on order of 64
CNUs for a single OLT port (give or take). The key, though, is that we need
to design the system to handle more than that.
Business services might be closer to the 64-128 number to start, ramping
down to smaller numbers over time.
Note that the above is my personal opinion; I have not talked to my members
to generate a specific consensus position on that topic.
Marek, as to your point about SNR variance? The problem is that even if
there are only 2 devices, while statistically the odds are lower that
there's a significant spread in SNR, there's still a chance that they will
be spread quite a bit. Additionally, if you're looking at multiple OCUs
aggregated to a single OLT port, the odds of significant SNR spread between
devices sharing that single OLT port (because they're on different OCUs) is
actually quite high.
I believe very strongly that we need to adopt some sort of mechanism in EPoC
that allows for different CNUs to operate with different Modulation and
Coding Schemes (MCSs). I believe you can get the majority of the benefit
with a limited number of "profiles" that CNUs could be grouped into, but I
think you will need at least on order of 3-4 of these.
Thanks.
Matt
From: <Salinger>, Jorge Salinger <Jorge_Salinger@xxxxxxxxxxxxxxxxx>
Reply-To: Jorge Salinger <Jorge_Salinger@xxxxxxxxxxxxxxxxx>
Date: Sunday, October 7, 2012 4:07 PM
To: "STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx" <STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx>
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Right, I meant to describe things in the longer term.
Maybe in a nutshell, on could say that networks will be designed in such a
way that the number of CNUs/OLT port will be to be similar to the number of
ONUs/OLT port, by aggregating several OCUs/OLT port initially as CNU
penetration is low, and by splitting OCUs into more OLT ports as the number
of CNUs grows. I think his would apply to both resi and business services.
Does that make sense?
Jorge
From: Edwin Mallette <edwin.mallette@xxxxxxxxx>
Date: Saturday, October 6, 2012 3:02 PM
To: "Salinger, Jorge" <Jorge_Salinger@xxxxxxxxxxxxxxxxx>, EPoC Task Force
<STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx>
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
In general it makes sense, I think what Jorge describes might be true in
time ? not sure that's true "day-1." In fact I can easily see a case where
we might dabble with aggregating as many CCDNs together to a single 10G-EPON
port to efficiently utilize the bidirectional capacity (and efficiently
utilize our Capex and Opex spend). In the early days where we have 120MHz
or less of downstream bandwidth, I can see attempting to aggregating 10
CCDNs together, as an example, for mostly residential services. As a result
you might see a pretty high number of scheduled services (LLIDs.)
Eventually I think we'll trend to having the number of CNUs in a CCDN
roughly equivalent to what we could see with ONUs per OLT port (32 or 64)
and accordingly we'd see the OCUs per OLT port trend to one or two if that
makes sense.
Ed
From: Jorge Salinger <jorge_salinger@xxxxxxxxxxxxxxxxx>
Reply-To: Jorge Salinger <jorge_salinger@xxxxxxxxxxxxxxxxx>
Date: Saturday, October 6, 2012 12:24 AM
To: <STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx>
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Personally, I think that for the use case of residential services, there
would be few OCUs/OLT port, yielding an effective number of CNUs/OLT port
that is similar as the number of ONUs/OLT port. For business services, since
the subscriber density is lower, there would be multiple OCUs/OLT port, and
even in that case the number of CNUs/OLT port would be lower than for
residential services. Does this make sense?
JD,
Ed,
Does what I describe make sense to you?
Thanks!
Jorge
From: Marek Hajduczenia <marek.hajduczenia@xxxxxxxxx>
Date: Friday, October 5, 2012 6:12 PM
To: "Salinger, Jorge" <Jorge_Salinger@xxxxxxxxxxxxxxxxx>, EPoC Task Force
<STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx>
Subject: RE: [802.3_EPOC] Questions on varanese_01_0912.pdf
Jorge,
Makes sense. Have you given a thought to bandwidth available per CNU in the
case of ?hundreds of CNUs per ONU? ? That seems that a single 10G OLT port
might be shared by several thousands CNUs, leaving bandwidth per CNU less
than attractive.
Just trying to see what the target scenario might be and what number of CNUs
to expected subtended to a single OLT port
Marek
From: Salinger, Jorge [mailto:Jorge_Salinger@xxxxxxxxxxxxxxxxx]
Sent: Friday, October 05, 2012 22:57
To: STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
I wanted to share some thoughts on this?
I think that there are two deployment scenarios and two use cases that MSOs
are considering. One of the deployment scenarios is where the OCU is
deployed close/next to the node and the EPoC signals traverse amplifiers,
and the second where the OCU is deployed past the amplifier. The two use
cases are business and residential services.
For the use case of residential services, in the first deployment scenario
there could be hundreds of CNUs per ONU, and in the second there would be
far fewer CNUs/OCU. For the use case of business services, in either of the
deployment scenarios there would be far fewer CNUs per ONU.
Does this make sense?
Thanks!
Jorge
From: Marek Hajduczenia <marek.hajduczenia@xxxxxx>
Reply-To: Marek Hajduczenia <marek.hajduczenia@xxxxxx>
Date: Friday, September 28, 2012 5:47 AM
To: EPoC Task Force <STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx>
Subject: Re: [802.3_EPOC] Questions on varanese_01_0912.pdf
Thank you Andrea,
Please see inline
Marek
From: Marek Hajduczenia [mailto:marek.hajduczenia@xxxxxx]
Sent: Friday, September 28, 2012 10:29
To: STDS-802-3-EPOC@xxxxxxxxxxxxxxxxx
Subject: [802.3_EPOC] Questions on varanese_01_0912.pdf
Dear colleagues,
Here are some questions on the varanese_01_0912.pdf presentation which did
not get sufficient time for discussion. I?d appreciate if they were answered
via reflector so that everybody benefits from these clarifications:
- How many taps have been examined in total in this study? I do not
care much about names, types but rather to see what the sample size we are
looking at and whether it is representative of a large network as a whole
rather than a single CMT port or not.
[AG] I think it is mentioned in the slides, we have 140 subscriber (CNU)
ports for the model, we considered all of them in the analysis. As you can
see from the curves, since this is a model and small variability have not
been included, SNR curves overlap when users are attached to same last
splitter ? in reality, the CDF is more continuous like shown in the measured
valued, rather than step-wise. Does that answer your question?
[mh0928] This begs a question then. Is this a number of CNUs that you?d
consider typical? What I am trying to understand whether the scenario that
was presented is the worst-case, best case or average (what can be expected
in majority of deployments)? Would it be possible for an operator to use
more tailored service groups to optimize them for SNR performance and to
avoid complicating the design of devices, allowing for more optimized
performance, rather than complicating the design of active devices?
- What is the most probably SNR distribution for a much smaller
population of CNUs connects to a single CLT port? I assume that you will see
some difference in SNR but it is not very likely to be as high as it was
presented at the meeting for the whole measured population of taps and
ports.
[AG] That may be the case, depends on how the few users are distributed in
the plant. However, the question is whether this would be representative of
a realistic plant ? measurements are over population of ~240 modems, so it
seemed to us that 140 (already smaller) was in the correct ballpark. Do you
see any use case for much smaller plants, we could include in the analysis?
[mh0928] This is something that operators should speak to. However, when I
look at the OLT driven deployment model with several CLTs deployed in field,
I?d not expect each CLT to be connected to 150+ CNUs. That would easily
reach thousands of CNUs visible to a single OLT, which brings the available
bandwidth down drastically, while burning a lot of bandwidth on scheduling
overhead. I?d like to understand the trade-off here, that is all.
Please consider presenting more focused study for the next meeting, focusing
on a number of drop sections to show what is expected to be seen on a single
CLT port. While I am not against adaptive loading on per CLT port, I do not
believe that this contribution has sufficient footing to justify adaptive
loading on per CNU basis.
[AG] What we shown is one CLT port and 140 CNU ports attached for the
modeled plant ? for measured values, each plant is one node and ~240 CM all
attached to the same coax distribution tree (Comcast may provide more
clarifications in case) ? we can make it more clear and improve in the next
steps.
Regards
Marek Hajduczenia, PhD
ZTE Portugal
Technology Strategy Department
Edifício Amoreiras Plaza,
Rua Carlos Alberto da Mota Pinto, nr. 9 - 6 A,
1070-374 Lisbon, Portugal
Office: +351 213 700 090
Fax: + 351 213 813 349
Mobile: +351 961 121 851 (Portugal)
_____
<="" p="">
_____
<="" p="">
_____
_____
<="" p="">
_____
<="" p="">
_____
<="" p="">
_____
_____
<="" p="">
_____
<="" p="">
_____
<="" p="">
_____
________________________________________________________________________
To unsubscribe from the STDS-802-3-EPOC list, click the following link:
https://listserv.ieee.org/cgi-bin/wa?SUBED1=STDS-802-3-EPOC&A=1