Re: [RE] Lastest white paper release and tomorrow's meeting
Thank you for your response. This goes a long way towards clarifying things
for me.
A constructive counter-proposal to the multicast stream transport is a
reasonable request. I'll try to find some time to work something up.
It seems like a couple big-ticket design elements (extra buffering, queue
sorting) are motivated by the analysis in the bursting and bunching annex. I
will definitely take another read through of that. Who else has reviewed it
in detail?
-----Original Message-----
From: owner-stds-802-3-re@ieee.org [mailto:owner-stds-802-3-re@ieee.org] On
Behalf Of David V James
Sent: Wednesday, July 06, 2005 8:20 PM
To: STDS-802-3-RE@listserv.ieee.org
Subject: Re: [RE] Lastest white paper release and tomorrow's meeting
Kevin,
I'll try for some quick answers right now, but we may want
to queue some topics for future discussion.
>> The paragraph following figure 5.8 doesn't make much sense to me.
It made sense to me, but I have the handicap of being the author.
Perhaps you could suggest modifications, or we could chat on the
phone to work out better wording?
>> Figure 5.20 shows FIFO's on the MII side of the PHY.
>> The main working FIFOs are typically on the MII side of the MAC.
>> Accurate receive packet time stamping can be obtained by
>> monitoring the MII receiver carrier sense (CRS) signal.
I'll have to defer to the PHY experts on this one. I was
simply trying to illustrate that _if_ significant FIFOs
are in the PHY (as they may be on generalized PHY-addressible
OIF interconnects), then accuracy is better if a signal
could be provided when the frame first arrived.
Do you have a preferred technology/implementation independent
wording/illustration?
>> In order to enforce bandwidth subscriptions, it seems
>> there needs to be an association between sourceID:plugID
>> and some sort of identifier in the stream data packets.
>> The network needs to identify these packets and associate
>> them with a stream, measure bandwidth consumed by the
>> stream and potentially drop packets if the subscription
>> terms are violated. Figure 5.22 and bullet (a) in
>> section 5.6.3 seem to indicate that a multicast destination
>> address for each stream is the preferred association.
>> We should be aware that this solution creates the following side effects:
>> 1/ All media data is always multicast.
>> 2/ ClassA "routing tables" must store a multicast destination
>> MAC (I'll call it destinationID) along with sourceID and plugID.
>> 3/ destinationIDs must be unique on the network. We'll
>> need some network-wide means of allocating destinationIDs.
>> We'll need to deal with destinatationID collisions when two
>> separate working networks are connected together.
On topic (1), this seemed to be a reasonable restriction.
That would allow smooth transitions from 1-to-N listeners.
On topics (2,3), your concerns are valid. The alternative of
using a denigrated G.1.1 solves these problems, but appeared
to have "too many changes to bridges" concerns.
Was there another option that you think should be considered
in more detail?
>> The pacing feature described in section 5.7 appears to
>> insert an isochronous cycle's worth of gating delay for
>> each switch hop. I assume the motivation here is to keep
>> traffic flowing smoothly.
Yes.
>> Many applications would prefer to have a fixed, longer
>> latency than to have latency dependent on network location.
Probably better to state this as "fixed, typically longer",
since the Annex illustrates that w/o this fixed delay,
bunching can happen and may actually generate a longer
worst case.
>> Also, holding gigabit data for 125us implies an additional
>> 120Kbits of buffer per port. There will be a cost associated with this.
Yes, there is a cost for this, but it is known to work.
And, the buffer may be smaller than those required in
a bunching-tolerant bridge.
If you have a preferred alternative, I'm sure everyone has
the flexibility to consider alternative approaches.
While I (and I suspect others) don't like the buffer requirements,
good alternatives haven't been forthcoming.
>> Transmission gating described in section 5.7.4 will only work
>> properly if ClassA packets in the transmission queue are
>> sorted according to isochronous cycle number. Are the
>> multiple arrows into the queue in figures 5.26b and 5.27b
>> trying to indicate the presence of sorting hardware?
Yes.
Due to receiver and transmitter cycle-slip, early and late
frames must have distinct precedences. This could be distinct
FIFOs, tag-ordered non-FIFO queues, linked-lists with multiple
heads, etc.. The arrows were intended to illustrate this.
Could this be described better?
Words or figures are most welcome.
>> I've taken in a first reading of Annex F. I would like to
>> understand how the graphs of latency vs. switch hops were
>> generated. These graphs show an unexpected exponential
>> relationship between switch hop count and latency.
The exponential relationship was no surprize to me, but it
took a while to figure out how to illustrate the conditions.
The base problem is that if each of N stations has a burst,
the the resulting bridge output effectively has to deal
with an N-1 length burst.
One can argue that this loading is not realistic, but that
was not the point. Guaranteed within any topology was the
constraint being evaluated/simulated.
These graphs were generated in an amazingly primitive fashion.
The previous time sequences were generated manually, using
(ugh) FrameMaker graphics, with the snap-grid on.
When generating the time sequences, an attempt was make to use
worst case collision conditions (although I can't guarantee
these are actually the worst). In some cases, contending
traffic was momentarily stopped, as though the variable-rate
traffic was dramatically reduced or was paused. It is a
bit ironic, that the worst case bunching can actually occur
when the offered load is reduced, but at just the wrong time!
This is _certainly_ not a typical scenario, but does represent
a scenario that could occur. Since we are trying to illustrate
worst case guaranteed, not typical expected, this seemed
to be a fair methodology.
The ticks on the time sequences were counted to generated
the "cycles" numbers in Table F.6 and others. Manual
processing used a Windows calculator to convert cycles
to time. Time values were plotted via visual placement.
This certinly wasn't he most elegant of simulations
(Ooohhhh, I feel so embarassed...), but is easy to verify
via visual observation and any '+-*/' capable calculator.
In a some cases, the behavior was exponential but the width
of the page was insufficient, so the dotted lines continue
beyond the last computed data points.
The point of a pretty document is to improve readability and
the best judge of readability is not the overly familiar
editor. So, specific "from-to" text proposals are always
helpful when I can't clearly visualize the difficulty.
As always, thanks for the most careful review. Feel free to
call if some of this doesn't make sense and an interactive
conversation would help.
Respectfully,
DVJ
-----Original Message-----
From: owner-stds-802-3-re@ieee.org [mailto:owner-stds-802-3-re@ieee.org]On
Behalf Of Gross, Kevin
Sent: Wednesday, July 06, 2005 3:35 PM
To: STDS-802-3-RE@listserv.ieee.org
Subject: Re: [RE] Lastest white paper release and tomorrow's meeting
I've had a chance to read a bit further into the working paper. I have a few
comments and question:
The paragraph following figure 5.8 doesn't make much sense to me.
Figure 5.20 shows FIFO's on the MII side of the PHY. The main working FIFOs
are typically on the MII side of the MAC. Accurate receive packet time
stamping can be obtained by monitoring the MII receiver carrier sense (CRS)
signal.
In order to enforce bandwidth subscriptions, it seems there needs to be an
association between sourceID:plugID and some sort of identifier in the
stream data packets. The network needs to identify these packets and
associate them with a stream, measure bandwidth consumed by the stream and
potentially drop packets if the subscription terms are violated. Figure 5.22
and bullet (a) in section 5.6.3 seem to indicate that a multicast
destination address for each stream is the preferred association. We should
be aware that this solution creates the following side effects:
1/ All media data is always multicast.
2/ ClassA "routing tables" must store a multicast destination MAC (I'll call
it destinationID) along with sourceID and plugID.
3/ destinationIDs must be unique on the network. We'll need some
network-wide means of allocating destinationIDs. We'll need to deal with
destinatationID collisions when two separate working networks are connected
together.
The pacing feature described in section 5.7 appears to insert an isochronous
cycle's worth of gating delay for each switch hop. I assume the motivation
here is to keep traffic flowing smoothly. Many applications would prefer to
have a fixed, longer latency than to have latency dependent on network
location. Also, holding gigabit data for 125us implies an additional
120Kbits of buffer per port. There will be a cost associated with this.
Transmission gating described in section 5.7.4 will only work properly if
ClassA packets in the transmission queue are sorted according to isochronous
cycle number. Are the multiple arrows into the queue in figures 5.26b and
5.27b trying to indicate the presence of sorting hardware?
I've taken in a first reading of Annex F. I would like to understand how the
graphs of latency vs. switch hops were generated. These graphs show an
unexpected exponential relationship between switch hop count and latency.
From: owner-stds-802-3-re@ieee.org [mailto:owner-stds-802-3-re@ieee.org] On
Behalf Of David V James
Sent: Tuesday, July 05, 2005 6:44 PM
To: STDS-802-3-RE@listserv.ieee.org
Subject: [RE] Lastest white paper release and tomorrow's meeting
All,
Based on last meeting's consensus, the white
paper now includes multicast-address-selected
classes, pacing for classA and shaping for
classB.
Things seemed to work out during the writing,
yielding a good compatibility classB as well
as achieveable low-latency bridge forwarding
possibilities:
1 cycles (125 us) for Gb-to-Gb/100Mb
2 cycles *250 us) for 100Mb-to-Gb/100Mb
I have placed these on my DVJ web site, but assume
that Michael will transfer them to the SG web
site soon. You may find them at:
http://dvjames.com/esync/dvjReNext2005Jul05.pdf
http://dvjames.com/esync/dvjReBars2005Jul05.pdf
If anyone has a chance to read the pacing stuff,
we can discuss this briefly at tomorrow's adhoc
meeting. In case anyone missed the announcement,
tomorrow's interest group meeting is as follows:
REsE interest group conference call, code 802373
Wednesday July 6, 2005
2:00 pm - 4:00 pm
Event Location: 1 877-827-6232 / 1 949-926-5900
Street: 3151 Zanker Rd
City, State, Zip: San Jose/CA/95134
Phone: 1 877-827-6232(US)/ 1 949-926-5900(non-US)
Notes:
RSVP mikejt@broadcom.com if you are calling in so I can confirm the number
of phone ports
I normally wouldn't want to discuss things on such
short notice, but I'm aware that Michael will be
there tomorrow and not a week later.
Cheers,
DVJ