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

RE: OFC revived; 2.5 Gbit Ethernet




Patrick,

I will try to answer some of your questions, see below.

jonathan

> -----Original Message-----
> From: owner-stds-802-3-hssg@xxxxxxxx
> [mailto:owner-stds-802-3-hssg@xxxxxxxx]On Behalf Of Patrick Gilliland
> Sent: Wednesday, March 01, 2000 10:01 PM
> To: stds-802-3-hssg@xxxxxxxx
> Subject: Re: OFC revived; 2.5 Gbit Ethernet
>
>
>
> Jonathan,
>
> I think your solution is a good one.  Maybe OFC
> has finally found a place in our lives if we choose
> to use it in a WDM multichannel system.
>
> However, every time I contemplate multichannel WDM
> systems I am confronted with a couple of nagging
> issues with which I still grapple:
>
> 1)  If we decide on a WDM standard, why would we not
> want to use it with the present 1.0 Gbit/s systems?

This decision, like most good engineering decisions, is part economics.
There are companies who have, or at least market, WDM systems for 1 gig.
Technical feasibility is not sufficient for the IEEE. There must be "broad
market potential" also. This means that a preponderance of the industry must
agree that it is not a niche market. If you think this is the case, then you
would get together with others who share that opinion and request a "call
for interest." If there is sufficient interest, a study group would be
created to develop the standard.

>
> 2)  If we go ahead with a 4 channel WDM system, it
> could conceivably operate at 4.0 Gbit/s with the current
> GbE standard components, save the VCSEL and WDM filters.

Yep.

>
> 3)  The WDM standard could be evolved separately from
> the next generation 10GbE serial standard.  This way,
> we could immediately upgrade the capacity of today's
> LAN and provide a path to 40 Gbit/s capacity when we
> finally reach agreement on a 10GbE standard protocol.
>

Yep. There will certainly be markets for aggregation of 10 Gb/s pipes. These
will provide good interim solutions to the 100 Gb/s Ethernet standard which
I expect in 2006. Given current growth, this will barely be in time.

> The way I understand the multichannel proposals currently
> in play is perhaps lacking.  I would welcome a different
> perspective.  Why bother splitting the data up into four
> channels, transmit them over different wavelengths, and
> then recombine them at the far end unless there is some
> compelling reason to do so?
>

Some people believe the compelling reason to do so is price/performance.
You've seen all the presentations. Right? You may disagree. Other people
have. What would your solution be for 100 meters over existing multimode
fiber be? That is one of the objectives the committee decided for itself.
Some of your customers really want that!

> The only reason I can supply is there must be some aspect
> of this data which binds it together logically.  The only
> one I can imagine might be the need to send wide data words
> at high rates from server to server.
>
> A modern processor might have a 1GHz clock and 32 bits
> in a word.  This would imply the usefulness of a 10Gb/s
> chip-chip interconnect.  Bus architectures commonly in use
> today such as PCI do not require 10Gb/s bandwidths.
>

This implies that you do not believe there is any value in aggregation
within a switch/router. OK, perhaps that is a bit harsh. But, you must admit
that your statement ignores the traditional arguments for higher bandwidth.
Even so, by the time the 10 GbE standard reaches the masses in large
volumes, we can expect that processors will have at least 4 GHz clocks with
32 bit paths. According to Amdahl's law, this would require I/O of 4 Gig *
32 / 10 = 10 Gig. Interesting, hey? Now, one could argue that I/O is not
networking. But, it sure is moving that way.

> I assume therefore, we must be preparing 10 GbE for a
> development in backplane technology such as Infiniband.

That will certainly also happen, over time.

>
> Some additional contradictions stir:
>
> 1)  Most network engineers from Ethernet backgrounds would
> not admit the driving force behind a network proposal is
> genetically a point-to-point backplane extension.
>

Right, network engineers think about networks. Backplane extensions will be
left to bus extension engineers (e.g. Infiniband). Expect 2.5, 10, and 40
Gb/s (encoded data rate) buses soon. This may use some of the same optical
hardware when distances require it.

> 2)  The parallel transfer of data from server-server
> sounds too much like Fibre Channel.  This statement
> might provoke true network engineers like Ethernet guys,
> but one cannot avoid the comparison to the early Fibre
> Channel goals.

Fibre Channel is already starting on 10 Gig FC. They will solve the SAN
problems. It will likely use the same PHY hardware.

>
> The alternative answer to why we need 10GbE WDM networks
> is perhaps to provide additional transmission capacity
> for 1GbE service over existing infrastructure.  I see
> this need very clearly as one which must be addressed
> in the short term.
>

While this may another reason, it is not the only reason. If this were true,
then one could argue that all that GbE really needed to be is a way of
providing transmission capacity for 10 Mb/s Ethernet.

> Errata:
>
> 1) Inasmuch as we choose to call a 4 channel WDM system
> operating at 2.5Gb/s a 10Gbit network, we may also term
> a 10 channel WDM network operating at the basic GbE
> 1.0Gbit/s rate a 10Gb network.  In the case of the latter,
> we might hesitate to term it 10GbE because we would not
> have to evolve any new standard for Ethernet protocol.
>

Sure, if you are willing to dedicate 10, 1 Gig MAC engines operating with
link aggregation. On the other hand, how will this run over the existing MMF
and SMF serial infrastructure?

> 2) Why is it we insist on terming 4 channel WDM systems
> at 2.5Gbit/s a 10GbE network?  Aren't we really working
> on 2.5 Gbit Ethernet with a 4 channel WDM overlay?  How
> could a 2.5X increase in the basic serial line rate not
> be considered an evolutionary improvement similar to the
> 2X Fibre Channel systems presently sampling?

Using this logic, 1000BASE-T, at 250 Mb/s over 4 pair of copper isn't a 1
GbE solution. I would think that the majority of systems integrators don't
much care what the PHY is doing to ship the bits. They look at the bandwidth
across the MAC interface. 1 gigabit per second. In the case of 10 gig, they
would expect 10 gigabits per second. That is what the standard does. If this
was accomplished with 1,000,000 "sub-space" parallel links, who cares. It's
10 gig.

>
> The above is especially true when one considers the genesis
> as a chip-chip or backplane interconnect.  I guess 2.5GbE
> doesn't sound as good as 10GbE.

For 2.5 GbE to be interesting, we would already have to be shipping it in
volume according to the Moore's law curve (in fact, we should have had it in
early 1999). That represents a level of churn that is a bit much. Don't you
think? In 1990 we saw 10 Mb/s Ethernet. In 1995, 100 Mb/s. In 1998, 1000
Mb/s. In 2002 10,000 Mb/s. This is approximately 10 X every 4 years. It
takes us about 3 years to build the standard. I, for one, just can't see
having multiple standards groups running in parallel so that we can complete
a standard every year (i.e. > 2X the speed every year).

>
> 3) What have we learned from attempts to use parallel data
> communications over distance (HIPPI, SCSI, etc.)?
>
> I think the answer to the above is parallel architectures
> do not scale well - especially when switching.
>

True. But at any point in time, parallel is always used when the practical
limitations of serial get in the way. Where this not true, all chips would
have single pin interfaces.

> Summary:
>
> So, in summary, I must agree with you there is a low burden
> Open Fiber Control solution available to us for WDM network
> proposals.  I also think we should develop WDM for the LAN
> environment independently from the protocol.
>

There is no reason why you should not do so. Other transceiver companies
have done so already. Please forgive the Ethernet standards community for
not wanting to standardize "every good idea." We are already worked pretty
hard doing what little we are doing.  :-)

> Best Regards,
>
> Pat Gilliland
> patgil@xxxxxxxxxxx
>
> -------------------------------------------------------------
>
> At 09:18 PM 2/26/00 -0600, you wrote:
> >
> >Patrick,
> >
> >Summary: Yes, open fiber control could simply (less than 50 gates) and
> >easily reduce transceiver costs for WDM, PAM, and Parallel
> transceivers (and
> >maybe even serial). Perhaps, now that the transceiver companies
> have figured
> >out how difficult the budgets are to achieve with low-cost optics, it is
> >time to address this in greater earnest.
> >
> >We have discussed this strategy a number of times at various
> HSSG meetings.
> >It was, in point of fact, first brought up by yours truly at the
> "call for
> >interest meeting" last March. It is certainly not popular with
> transciever
> >companies. Even so, personally, I don't think that people are being quite
> >fair here. One does not need to utilize timing at all to make "OFC" work.
> >Each of the colors or fibers (i.e., each laser) can be designed to be
> >inherently laser safe (NOFC-like). You just can't turn on more than one
> >laser in the multi-laser system until the link is closed. It is
> very simple
> >logic to design. Dead simple. For a four laser system, you potentially
> >increase the optical power usable by 4X or 6 dBm. Since timing
> would not be
> >used, distance and latency would not enter into the equation (even for
> >multihop systems like FC-Loop; a 10 Gig FC-loop! Yikes!).
> >
> >In the case of PAM-X, the Rx must be able to correctly detect the lowest
> >level signal. This lowest level signal could therefore easily be used to
> >check for a closed link. The system would simply not use the higher power
> >signals required by the coding until the closed link was
> detected. This, for
> >PAM-5, would also allow a 4X or 6 dBm effective increase in power in the
> >link.
> >
> >The funny thing is, if the SERIAL transceivers used this same
> trick (yep it
> >would probably add a Mux and resistor divider to the DC control
> circuit on
> >the laser and modify the sense circuit on the light detection
> circuit, which
> >is likely needed anyway due to the long run lengths of zero's being
> >proposed), even the serial guys could fix some power budget problems.
> >
> >Like with the original OFC based system, this additional power and power
> >budget can be traded off AGAINST higher sensitivity and tighter
> >manufacturing tolerances, thereby reducing the cost of the
> transceiver (the
> >original goal for FC-OFC). While it is true that FC threw out FC-OFC, my
> >recollection was that this was primarily due to the TIMING problems
> >encountered running FC-Loop over optics, which is not a problem here
> >(remember, no timing). It is also the case that there were companies that
> >did not meet the timing specifications because they forgot to include the
> >turn-on and sense delays of the analog portion of the laser's DC control
> >circuit to the overall timing equation. After shipping product
> this way for
> >a fair period of time, these companies asked the FC group to go back and
> >modify the timing numbers to allow them to "meet spec." The
> problem was not
> >with the original specification or the concept, but because we
> had to find a
> >set of numbers that several companies could all say they met.
> Not being able
> >to find any, we had to revamp the entire definition to prove
> that equipment
> >in the field would continue to work. It does. The last issue
> with FC's OFC
> >was the 10 second max time delay for turn on. This was, again,
> an artifact
> >of the use of timing and, again, not an issue here.
> >
> >None of these issues apply to the scheme above (and below).
> >
> >I believe that some WDM advocates do not like the idea of using OFC (this
> >non-timed version) because the losses in the Tx optical system (laser to
> >fiber path) are sufficiently severe that they can't run the lasers any
> >hotter anyway. For some of the parallel fiber people, some interesting
> >mathematical games are played in order to argue that the light from the
> >multiple fibers don't quite add.... Both of these seem to be rather
> >short-sighted positions.
> >
> >jonathan
> >
> >> -----Original Message-----
> >> From: owner-stds-802-3-hssg@xxxxxxxx
> >> [mailto:owner-stds-802-3-hssg@xxxxxxxx]On Behalf Of Patrick Gilliland
> >> Sent: Saturday, February 26, 2000 4:46 PM
> >> To: stds-802-3-hssg@xxxxxxxx
> >> Cc: kardontchik.jaime@xxxxxxxxxxx
> >> Subject: Re: open fiber control in PAM-5
> >>
> >>
> >>
> >> Jaime,
> >>
> >> No doubt there are schemes to overcome any
> >> difficulty such as the Class 1 power limitations
> >> imposed by the IEC and CDRH.
> >>
> >> The option of using Open Fiber Control does not
> >> seem to be particularly attractive.  During the
> >> FC-0 days of Fibre Channel it was used to overcome
> >> the same Class 1 power limitations.
> >>
> >> A great deal of time and effort was spent trying to
> >> create an OFC protocol which would guarantee multi-
> >> vendor interoperability.  We even had an OFC working
> >> group to address these issues.  The FC-0 equipment
> >> which is still in service today is in many cases, not
> >> interoperable because of OFC timing incompatibilities.
> >>
> >> This is especially true at longer distances, as the
> >> latency requires longer and longer OFC related outages
> >> as the interested parties wait for the other partner to
> >> respond with the correct sequence of pulses before they
> >> initiate full duty cycle transmission.  I believe Fibre
> >> Channel correctly eliminated an obstacle to interoperable
> >> systems in FC-1.  Sometimes simple is better.
> >>
> >> Best Regards,
> >>
> >> Pat Gilliland
> >> patgil@xxxxxxxxxxx
> >>
> >> -----------------------------------------------------
> >>
> >>
> >> At 02:55 PM 2/22/00 -0800, you wrote:
> >> >
> >> >Hello 10G'ers,
> >> >
> >> >I would like to eliminate another misconception
> >> >regarding the serial at 5 Gbaud vs the 4-WDM
> >> >at 1.25 Gbaud approaches in PAM-5: the supposed
> >> >signal power advantage of the serial approach
> >> >due to eye-safety limits.
> >> >
> >> >I will show below that the launched power
> >> >PER CHANNEL in 4-WDM can be safely set at the
> >> >same level as the total launched power in the
> >> >serial approach, due to the redundancy in
> >> >its receiver (no single point of failure).
> >> >
> >> >   ---> Since the launched power level per
> >> >        transmitter of the serial and 4-WDM
> >> >        approaches will be the same, 4-WDM will
> >> >        enjoy a 12 dB advantage in SNR, due to
> >> >        its receiver bandwidth being 4 times
> >> >        smaller (same signal power, much smaller
> >> >        noise power).
> >> >
> >> >The key is in the "open fiber control" method.
> >> >How do I envision this method in the PAM-5
> >> >specific case ?
> >> >
> >> >In a serial approach using, for example, 850 nm
> >> >lasers, the maximum launched power is -4 dBm.
> >> >
> >> >In a 4-WDM approach we could use the following
> >> >procedure to keep the launched power PER CHANNEL
> >> >at -4 dBm (for a total launched power of +2 dBm)
> >> >and still ensure a safe-eye environment:
> >> >
> >> >1) When a near-end node is connected to a link
> >> >and powered-up it will transmit a signal at -4 dBm
> >> >using only ONE transmitter and keep the other
> >> >three transmitters off. In this way the eye-safety
> >> >limit is satisfied.
> >> >
> >> >2) The near-end node will remain in this state
> >> >for as long as it does not sense a signal in
> >> >any of its four receivers.
> >> >
> >> >3) Also the far-end node, when it is powered-on,
> >> >will do the same: transmit on only one channel
> >> >using the maximum -4 dBm allowed for eye-safety
> >> >considerations.
> >> >
> >> >   It makes sense to use only one transmitter
> >> >   after power-up to send a life signal to a
> >> >   potential partner on the other side of the
> >> >   link, just to conserve power consumption
> >> >   as long as there is no answer from the other
> >> >   side.
> >> >
> >> >   Which transmitter should be used for sending
> >> >   this life signal ? For reasons that will
> >> >   become obvious later, the best choice is
> >> >   the transmitter that sends the PAM-5 encoded
> >> >   TA symbols, following the nomenclature of
> >> >   the 1000BASE-T standard (*)
> >> >
> >> >   For the following, remember that each
> >> >   receiver has four channels, named RA, RB,
> >> >   RC, RD.
> >> >
> >> >4) If any of the two partners senses and recognizes
> >> >   a signal in its RA-receiver, it will go
> >> >   to the next state of its state machine: it
> >> >   will switch-on the other three transmitters and
> >> >   begin transmitting IDLES, with each transmitter
> >> >   using a full -4 dBm launched power.
> >> >
> >> >   The advantage of using only the TA-transmitter
> >> >   during the first step of establishing a link
> >> >   becomes now obvious: in the 1000BASE-T,
> >> >   the RA-receiver has the capability to
> >> >   synchronize the receiver descrambler to
> >> >   the transmitter scrambler, by just using
> >> >   only the information embedded in the
> >> >   transmitted TA-symbols.
> >> >
> >> >   Hence, there is no danger that the receiver
> >> >   will confuse a spurious signal with the real
> >> >   signal: it has to be able to synchronize
> >> >   its descrambler and verify that the
> >> >   synchronization is indeed correct using just
> >> >   the embedded information in the transmitted
> >> >   TA-symbols, in order to make a positive
> >> >   identification. Without this identification
> >> >   it will not switch 'on' the other three
> >> >   transmitters.
> >> >
> >> >5) During normal operation, if any partner suddenly
> >> >   ceases to receive signals on ANY of its four receivers
> >> >   (that are tuned to four different wavelengths)
> >> >   during more than, say, 1 millisecond, it will switch
> >> >   back to its previous state, that is, shut off
> >> >   three transmitters and send a life signal
> >> >   using only one transmitter at - 4 dBm. This
> >> >   might be the case, for instance, when a
> >> >   technician opens the link at any point
> >> >   between the two partners.
> >> >
> >> >   (the loss of signal is easier to detect than
> >> >   the existence and validation of a real signal,
> >> >   hence, a very simple and robust no-signal-detect
> >> >   circuit may be used).
> >> >
> >> >   Notice that the four receivers, that are tuned
> >> >   to four different wavelengths, must malfunction
> >> >   in order to miss the "open link" event. It is
> >> >   this redundancy in a 4-WDM system that allows
> >> >   the use of a total + 2 dBm launched power during
> >> >   normal operation.
> >> >
> >> >The above procedure could also be a replacement of the
> >> >PHY Control State Diagram, Figure 40-15, of the
> >> >1000BASE-T standard (with further details to be added
> >> >later), since in 10 GbE we do not need the concept of
> >> >"master" and "slave" and loop timing used in 1 GbE
> >> >(that was used there to eliminate the Echo and NEXT
> >> >interferers).
> >> >
> >> >(*) during normal operation, after the link has been
> >> >    established, a transceiver sends through its four
> >> >transmitters quartets of PAM-5 symbols, {TA,TB,TC,TD},
> >> >with TA = {+2,+1,0,-1,-2} and similar for TB,TC,TD.
> >> >
> >> >Jaime
> >> >
> >> >Jaime E. Kardontchik
> >> >Micro Linear
> >> >San Jose, CA 95131
> >> >email: kardontchik.jaime@xxxxxxxxxxx
> >> >
> >> >
> >> >
> >> >
> >>
> >
>