Mike and Jugnu
You both have some very good points. On one hand we have the high
performance computing demanding highest bandwidth
achievable but on the other hand Jugnu brings up the fundamental
requirement for having successful Ethernet standard "mass market
potential".
Mike also mentions this project will take 3.5-5 years, IEEE project
traditionally have not taken this long this means then we are
too early. For example when 802.3ae was started in the 1999 10Gig
lasers/modulators existed for more than
15 years prior.
I have listed dilemma we are facing:
- Implementing 100 Gig in the near term means Nx10Gig
- Implementing 100Gig in few years the right answer might be nx25Gig
- Carrier want to leverage their existing DWDM layer which mean
baudrate in the 9.95-12.5 Gig
- If LAG implemented why not allow n to be 4?
- Operation with different width
- Backward compatibility XAUI, LX4 ?
- Greatest bandwidth demands (100+Gig) are on VSR links <50 m
but the longer reach >10Km
may be able to live with 4x10Gig.
All these means we should either define some sort of scalable
architecture or just define LAG method and
do not define any PMDs!
Thanks,
Ali
Jugnu,
OJHA,JUGNU wrote:
Mike,
It seems
more reasonable to me to consider
finally decoupling the physical pipe size from the rigid hierarchy used
in the
past. Why not simply define a scalable interface that allows inverse
multiplexing (physical layer aggregation – not the type of aggregation
you have described, which sounds like the current LAG) of an arbitrary
(within
some bounds, obviously) number of physical channels (10G) into a single
logical
link? The SONET/SDH and Digital Wrapper/OTN world already has
mechanisms
to do this (VCAT, LCAS), and dynamically to boot.
This would
provide a much more flexible,
scalable solution to customers.
On the surface, it seems to me that with flexibility comes complexity
which leads to higher cost. It's also not clear to me how the physical
layer aggregation you propose translates to a port on a switch. Would
I have to buy an N x 10G transceiver? Would it be a WDM-like
transceiver? Would this work with a single fiber-pair or
multi-strand? What would the relative incremental cost be (in
percentages, non monetary units) to scale up? Also, are you proposing
that this would scale beyond 100G? If so, how far? You mention
boundaries - I'm curious what you think the upper bound would be. I
hope you're planning to present something at the interim as it would
help me understand what you're really proposing and how that compares
to other ideas.
Regards,
Mike
In
particular, it would allow them to grow
capacity on any given link as needed, instead of having to install
10x10G
channels up front. Further, when they hit 100G, they wouldn’t be
stuck until some other solution is defined – they could continue to
grow.
Respectfully,
Jugnu Ojha
Avago
Technologies
John, et al.,
>During our first
meeting, I anticipate spending a lot
of time focusing on objectives. At the >closing
plenary I
highlighted two issues / objectives that the SG would have to consider:
>
> Tradition of
10x leap in speed
I
think
the speed
increase has to be 10x. The standards development process will take at
least 3.5 to 4 years to complete. Anything less than 100G will force
people who are currently aggregating 10G links to continue to use
aggregation,
only using fewer higher-speed, and more expensive links. End users
prefer using a single link over aggregating physical-layer links into a
logical
link because of the complications that come with aggregation. The data
in
the CFI presentation was just a sample of cases in which network
operators
we're aggregating 10G links to accommodate the demand on their
networks.
There will be many more by 2011 (when I expect there would be 'real'
products
on the market).
> Multiple
Reach Targets
>
It was also presented that the focus of this effort wasn’t for a
desktop
application, and >that the cost model needs to be considered.
I believe we need to adjust the cost model in such a way that it is
aligned
with the ecosystem. It is unreasonable, in my opinion, to expect a
10x/3x
model to apply to systems designed for wide-area/metro-area networks.
I
also think it's short-sighted to ignore the rest of the ecosystem and
develop
Ethernet only in the part of the ecosystem where the original cost
model
applies.
Regards,
Mike
--
Michael J. Bennett
Sr. Network Engineer
LBLnet Services Group
Lawrence Berkeley Laboratory
Tel. 510.486.7913
|
|