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

RE: Data rate standards vs internal switching standards




Bob:

Though not all switches use shared memory, all packet data switches do use
managed buffer memories to control congestion. These buffers may be input,
output, or shared but are always present and are required for TCP and other
data protocols.

As to clocking for packet data switches, the backplane, memory, or fabric
used in the core of a switch is always asynchronous to receive data. A
major function of the MAC chips (not the architectural MAC) is to make a
clock domain conversion between the switch core clock and the line clock.
MAC chips also use external transmit oscillators which are asynchronous to
the switch core.  

Since MAC chips perform the clock domain conversion between the line clocks
and the switch core clock there is no advantage in using an even 10.000
clock rate for the line clock. The clock will go through a clock domain
conversion anyway. On the other hand there is a great disadvantage in
having a PHY clock which is asynchronous with the MAC receive and transmit
clocks. Such a design forces an asynchronous clock boundary into the PHY
which otherwise would not be necessary. 

The clock rate of the MAC should be synchronized with the PHY so the only
clock domain conversion which needs to be done is the one already required
in the MAC. The MAC transmit and relieve clock rates therefore should
depend on the encoding system. For the scrambler system which I've proposed
the ideal MAC clock rate would be 9.953280 G with byte by byte HOLD
limiting the delivered data rate at the 10GMII to 9.584640 G bps.

I am opposed to having the rate set as a management parameter or negotiated
since these only add complexity to systems with no benefit to customers.

Paul

At 10:33 AM 8/4/99 -0700, Grow, Bob wrote:
>
>Roy:
>
>In your second paragraph below, you make some invalid assumptions about
>switches, and I find your arguements without substance.  Many switches are
>not shared memory. Shared memory has basically reached its limits for high
>performance fabrics.  (Memory speeds are not tracking the increase in
>throughput requirements and you don't benchmark well if you go wider than a
>minimum ethernet frame).  Alternate architectures do not share the data
>buffering for the "overall switch".  A small FIFO or even a PHY frame buffer
>is insignificant in monitoring the "overall 'health'" of the switch (just
>look at the per port buffering in gigabit switch specifications).  One of
>the things you try to avoid in a switch fabric design is allowing one port
>to consume too much of the buffer because if you don't it will eventually
>effect uncongested ports.
>
>I also think you are wrong is assuming that because a factor of 10 is less
>optimal than 2^n, it doesn't matter if the data rate is not an integer
>multiple.  I was using the same techniques you advocate for managing the
>different clock domains more than 20 years ago, and have also designed
>switches with both integer multiple clock and non-integer multiple clock
>domains.  It is because of this experience that I believe a 10.000 Gb/s MAC
>is the better system solution.  including some with similar clock domains to
>what you and because I know it I appreciate the simplification that an
>integer multiple gives in building a switch.
>
>--Bob Grow
>
>
>-----Original Message-----
>From: Roy Bynum [mailto:RBYNUM/0004245935@MCIMAIL.COM]
>Sent: Wednesday, August 04, 1999 7:19 AM
>To: Ariel Hendel; stds-802-3-hssg
>Subject: Re: Data rate standards vs internal switching standards
>
>
>
>Ariel,
>
>Actually, I am not in favor of a programmable IPG. I think that the IPG
>should be set to minimum for all frames in full duplex 10GbE. With 400
>bytes as the current average size of Internet 802.3 frames, I don't
>think that there will be enough "slop" to make up the difference
>between a 10.0 gb MAC and a 9.584 gb PHY. In the future, with more
>and more video based applications, the average size of the data frame
>will be increasing. This will only cause the MAC buffer discard rate
>to increase if the MAC and PHY are not data rate matched. I would much
>rather see the data rate be defined at the MAC, not the PHY. 
>
>I would much rather see the data buffered internal to the switch
>matrix than at the MAC. This will allow the overall switch to act as a
>buffer instead of perhaps only one output link that is operationally
>overloaded. It will also help network management to be able to monitor
>the overall "health" of a switch or network architecture without
>detrimental effect to the users. Since there are no standards,
>requirements, or optimal internal data interchange clock rates other
>than modulo 2, it does should not matter that the higher data rate
>interfaces are operating at a slightly different output clock. The
>technology to move data at inconsistent transfer rates through a
>system and between interfaces was invented almost 20 years ago. This
>is not something new.
>
>I am more concerned with how this technology will be implemented by
>the customers'. I am concerned with how relatively "lower tech"
>implemented and support people will be able use 10GbE. I am concerned
>about designers and implementers that do not, can not, or will not
>understand how users and their applications will make use of the
>extended LAN/MAN/WAN data networking environment. People will no
>longer be building extended data networks using routed meshed and
>semi-meshed virtual circuits, but will be using switched virtual
>segments over meshed semi-meshed 10GbE links. This is a very different
>implementation environment from the simple switched LAN that 100BT
>exists in. GbE has started to function in this realm. 10GbE will
>definitely be used like this in a major way.
>
>Thank you,
>Roy Bynum
>MCI WorldCom
>
>
>
>
>
>
>Date:     Tue Aug 03, 1999  4:28 pm  CST
>Source-Date: Tue, 03 Aug 1999 15:28:37 -0700 (PDT)
>From:     Ariel Hendel
>          EMS: INTERNET / MCI ID: 376-5414
>          MBX: Ariel.Hendel@eng.sun.com
> 
>TO:       Ariel.Hendel
>          EMS: INTERNET / MCI ID: 376-5414
>          MBX: Ariel.Hendel@eng.sun.com
>TO:       stds-802-3-hssg
>          EMS: INTERNET / MCI ID: 376-5414
>          MBX: stds-802-3-hssg@ieee.org
>TO:     * ROY BYNUM / MCI ID: 424-5935
>Subject:  Re: Data rate standards vs internal switching standards
>Message-Id: 99080322284244/INTERNETGWDN3IG
>Source-Msg-Id: <199908032229.PAA00067@mpk02.eng.sun.com>
>U-X-Mailer: dtmail 1.3.0 @(#)CDE Version 1.3.2 SunOS 5.7 sun4u sparc
>U-Content-MD5: NboaVGdw7MetJUdUqIoauA==
> 
>Roy,
>
>
>Just a quick reappearance to address your follow up questions.
>
>
>> 
>> I would like an additional clarification. In recognizing that the data
>> clocking standards are at the exposed interface of the data link, does
>> this mean that the standard applies to the MII between the MAC layer
>> and the PHY or does it apply more to the PHY?
>> 
>
>You have examples of both. The MII/GMII case is obvious.
>
>Other exposed interfaces were defined within the PHY to reflect real
>life partitioning into:
>
>- Clock recovery being the realm of exotic circuits designed by long
>haired gurus that seldom show up for work before 11 A.M.
>
>- The coding layer, can go into CMOS based MAC ASICs, and can be
>designed at any time of the day by many mortals skilled in the art of
>digital design.
>
>...
>
>> As for the OC rate standard, there are several standards for mapping
>> data into SONET/SDH transports. From an 802.3 view point, the
>> SONET/SDH standards can be treated as layer 1 functional processes.
>> From the other side of that argument, the current packet over SONET
>> (POS) standard for mapping required an additional standard for
>> inserting a layer 2 functionality between the layer 3 IP protocol and
>> the layer 1 SONET protocol. 802.3 does not have that requirement for
>> an additional functionality. In many ways it is more of a question of
>> how much of the SONET/SDH standards would not used for 10GbE,
>> depending on the implementation of the interfaces. 
>
>My emphasis was that the optimization of the data rate to 9.xyz is
>specific for a particular mapping and cannot be adopted in isolation
>unless such mapping is also embraced.
>
>Finally, I hope that your lack of objection to a 10Gbps rate with a
>statically programmable IPG is somehow a sign of agreement.
>
>
>> 
>> Thank you,
>> Roy Bynum
>> MCI WorldCom
>> 
>> 
>
>Regards,
>
>
>Ariel Hendel 
>Sun Microsystems
>
>
Paul A. Bottorff, Director Switching Architecture
Bay Architecture Laboratory
Nortel Networks, Inc.
4401 Great America Parkway
Santa Clara, CA 95052-8185
Tel: 408 495 3365 Fax: 408 495 1299 ESN: 265 3365
email: pbottorf@NortelNetworks.com