Hi,
----
> > 1/ I wonder why Section 4.2.1 does not include any normative statements
> on how
> > to handle the maximum character transmission rate ('cps' attribute). RFC
> 4103
> > states that "In receipt of this parameter, devices MUST adhere to the
> request
> > by transmitting characters at a rate at or below the specified <integer>
> > value." Isn't a similar statement needed in this document?
>
> The assumption has been that the associated procedures in 4103 apply.
> Note that section 6 in RFC 4103 continues:
>
> " Note that this parameter was not defined in
> https://tools.ietf.org/html/rfc2793 [https://tools.ietf.org/html/rfc4103#ref-16].
> Therefore implementations of the text/t140 format may be in use that
> do not recognize and act according to this parameter. Therefore,
> receivers of text/t140 MUST be designed so they can handle temporary
> reception of characters at a higher rate than this parameter
> specifies. As a result malfunction due to buffer overflow is avoided
> for text conversation with human input. "
>
> This note may be historic now, but for the T140-usage draft we may have the
> similar case when an implementation has not succeeded
> to implement support for the CPS parameter, or for dcsa at all.
>
> [MS] I have also read that paragraph, but to me it sounded as a historic
> transition problem. Also, as far as I read this text in 4103, the
> MUST is only about the design of the implementation, not about its actual
> operation. For instance, reception at a higher rate could
> be an optional feature turned off by default may not violate that MUST.
>
> We had specific wording for that case earlier, but I think we deleted most of it.
> Do you think we should insert a similar precaution in the t140-usage draft, but
> referring to other reasons than RFC 2793 interop?
>
> [MS] I don’t understand why the draft specifies a parameter if implementations
> shall just ignore the value. I guess the actual
> use of this parameter could be better specified in the document.
>
> [CH] I don't think we are saying that one shall ignore the value. We are simply
> realizing the fact that there may be legacy implementations that don't support
> the value.
>
> I don't understand that rationale. The WebRTC channel apparently requires a new implementation on both
> endpoints, i.e., there should be no legacy devices that don't support the value in that case. So, if a receiver following
> this spec decides to announce a cps value, I don't understand why any WebRTC sender should be allowed to send
> faster. If that was allowed, I would not understand why the parameter is even defined. And it is not clear to me for
> what the parameter would then actually be used.
You can implement T.140 data channels with current WebRTC stacks. They do support data channels.
> Also, if a non-WebRTC legacy device connected via a gateway ignores the value, then this is probably something to that requires attention in the gateway (e.g., by dropping text there).
We should not drop the parameter just because there might be legacy endpoint that doesn't support it. Sure, implementations need to be aware that such implementations exist, but eventually the legacy devices will hopefully be replaced or updated.
>> In any case, I don't think this is a real-life problem - in cases where T.140 is
>> used as intended. As Gunner mentioned, T.140 is not designed or intended for
>> transport of large amount of copy-pasted text but for human interactions.
>
> Then probably a big warning sign should be added to the document.
We do reference T.140, and I assume the scope and usage are described there. RFC 4103 does not have such statement either.
But, sure, if people think it is useful we can add a note saying that T.140 is intended for real-time interaction, not for transfer of text files, documents etc.
-----
> > 2/ Also, it is not really clear from the document what would happen if a
> peer
> > exceeds this maximum character transmission rate (or the rate allowed by
> > congestion/flow control). What happens if the sender types faster than the
> > 'cps' attribute (say, an automated chat bot)? I guess characters would be
> > dropped at the sender? In that case, no missing text markers would be
> displayed
> > in the receiver, right?
>
> I assume it could result in a buffer overflow sooner or later, but I think it is a
> local implementation issue how it is dealt with by the sender application.
>
> Perhaps Gunnar knows more about how implementations handle this?
>
> What the CPS parameter tries to prevent is buffer overflow in slow ancient
> receiving devices.
> Modern implementations will likely have buffer space enough for storing quite
> a large volume of text for eventual transmission to the limited device. It is
> instead another risk that appears. If you have a receiving device only able to
> present 4 characters per second which is the reality if you have interop with the
> old TTYs through a gateway, then a modern device user happening to paste a
> large piece of text or use speech to text technology will generate text in buffers
> that will take a very long time to present at the receiving device. The sense of a
> real-time conversation will be lost. (1000 characters will take 4 minutes to
> present!)
> The users need to realize that this technology is intended for human
> conversation. It is wise if implementations have the possibility for pasting in
> short texts, but it should not be used for document transfer.
>
> [MS] The document could better describe this, IMHO. As the data channel is
> different and has other capabilities, it is not clear to me that all limitations
> explained in RFC 4103 apply here as well.
> So, yes, it is a local implementation issue.
> [MS] From a transport protocol design perspective, the „cps“ parameter may
> realize some sort of (simple) flow control. And in that case sender and receiver
> typically have to agree on the semantics. And your example actually describes
> a typical flow control problem: If a 4 cps receiver receives 1000 characters, it
> may actually be interested in slowing down the sender. Could it do so, for
> instance, by sending cps=0? As far as I understand the I-D, this would be
> allowed. How to deal with that case is not a local implementation issue only.
> Well, we are here probably discussing corner cases of corner cases. But as far
> as I can see, the document does not comprehensively describe the use of the
> cps attribute.
>
> [CH] If the receiver cannot handle the received characters, in my opinion it is
> better if it uses the direction attributes (4.2.3) to indicate that it is not willing to
> receive text. I guess we could add some text about that. Something like:
>
> "If the receiver receives text at a higher rate than it can handle, e.g.,
> because the sender does not support the cps parameter, the receiver
> can indicate to the sender that it is not willing to receive more text using
> the direction attributes [ref-to-section-4.2.3]"
>
> With that being allowed, it becomes even less apparent to me why the cps parameter is needed at all.
It is used in order to help to prevent buffer overflow and sending of I-do-not-want-more-text requests to be sent.
I really don't understand what the problem is by having a way to indicate the maximum rate. Having legacy equipment that do not support certain protocol elements is nothing new in IETF, and not specific to data channels, but that hasn't prevented us from adding new features etc when we specify new versions of protocols etc.
> Also, SCTP has flow control, which should also be able to slow down a sender, or even stop it in case of a slow receiver. From the document is not
> really clear to me if the T.140-specific rate mechanisms (cps) and buffer timeouts (300/500ms) interact with underlying send and receive buffers for the data channel, and if so, how.
No matter what data channel usage you have, whatever you send and at whatever rate, the underlying transport mechanism (SCTP) obviously need to support it.
> I am not really familiar with WebRTC internals, but running some sort of rate-controlled traffic with soft real-time delay requirements over a reliable
> transport channel is typically not trivial. In this specific use case, the data rate is expected to be so small that it should usually work well. But, nonetheless,
> there could be corner cases in particular in challenging network environments and the document is silent about them.
The current version of the data channel specification does not contain load balance and rate control - applications need to make sure they don't try to send more than the data channel can handle.
However, I don't overloading the SCTP association is going to be a problem in the case of T.140.
---
> > 3/ Section 5.3. "Data Buffering" includes the following statement: "As
> > described in [T140], buffering can be used to reduce overhead, with the
> maximum
> > buffering time being 500 ms. It can also be used for staying within the
> > maximum character transmission rate (Section 4.2), if such has been
> provided by
> > the peer." I don't understand the second sentence. At first sight, enforcing
> > the 'cps' attribute does not only require a buffer, but also some sort of rate
> > shaper/policer (e.g., token bucket or the like). Do I miss something?
>
> The 2nd sentence talks about the case when the user input rate is faster than
> the maximum transmission rate (see question #2).
>
> Yes, so in the second case, the transmission procedure will detect that not all
> buffered characters are allowed to be transmitted, when the transmission
> interval (normally 300 ms) has passed. So they will be kept in a buffer at the
> sending side. Is that unclear so we need to clarify it?
>
> [MS] I had to partly reverse-engineer the handling of the buffer from RFC 4103,
> albeit it is not exactly clear if all considerations therein also apply to this
> specification. For instance, I believe that your additional explanation could be
> added to the document to better clarify the intention. And just to illustrate that
> there could be more ambiguity: To me, Section 5.3 could also imply that
> characters queued in the sender are dropped from the buffer if they cannot be
> transmitted within 500ms. That is a relatively small threshold for links with RTT
> of 200ms or more (e.g., satellite or slow GPRS links), e.g. if retransmissions are
> needed due to packet loss or congestion/flow control kicks in. Unlike RFC 4103,
> transport is now reliable and this could have some side effects even if the data
> rates in text conversations are very small. Out of my head, I would assume that
> a buffer threshold of 500ms will probably work very well in most parts of the
> Internet. But there are some areas with bad connectivity, high packet loss
> rates, etc. I am not sure of the proposed parameters really work well in such
> corner cases. Typically, such Questions would be answered by measurement
> data from running code.
>
> [CH] I don't know whether it is within the scope of this document to normatively
> define that characters shall be dropped if they cannot be transmitted within
> 500ms. In my opinion that is part of the generic T.140 procedures, not specific
> to data channels. But, perhaps we could say something like:
>
> "If a character cannot be transmitted within 500ms, and the sender
> chooses to drop the character from
> the buffer. If that happens, the sender SHOULD inform the user about it."
>
> That makes sense to me.
Thanks! I will add it.
---
> > 4/ Also in Section 5.3 is written: "An implementation needs to take the user
> > requirements for smooth flow and low latency in real-time text
> conversation
> > into consideration when assigning a buffer time. It is RECOMMENDED to
> use the
> > default transmission interval of 300 milliseconds [RFC4103], or lower, for
> > T.140 data channels". What is meant here by "or lower"? Does the
> document want
> > to recommend values much smaller than 300 ms, say, 1 ms? As explained
> in RFC
> > 4103, this could increase the overhead and bitrate, right? The absolute rate
> > values are relatively small for large parts of today's Internet, but couldn't
> > this text conversation be particularly useful in scenarios with very small
> > capacity of links (i.e., kbps range)?
>
> I suggest to remove the "or lower" part, since the recommendation is to use
> 300.
>
> Modern applications, especially speech-to-text are better at lower delays. So,
> having a 300 ms delay may be on the high side. Transmission intervals down to
> 100 ms will be experienced as an improvement for some applications. The load
> in both bandwidth and packets per second is still low compared to what audio
> and video (often used in the same sessions) cause.
>
> [MS] Indeed, bandwidth and packet rate will be small compared to other
> traffic. Nonetheless, this I-D somewhat specifies a service with real-time
> requirement that runs over a reliable transport channel, which could e.g.
> traverse highly congested links. Just assuming „the transport will just always
> work“ is a bit dangerous, IMHO.
> It is a RECOMMENDATION, so, yes, "or lower" could be deleted, but I prefer to
> leave it.
> [MS] Maybe the document could write something like „It is RECOMMENDED to
> use the default transmission interval of 300 milliseconds [RFC4103] for T.140
> data channels. Implementers MAY also use lower values". However, note that
> RFC 4103 backs the parameter of 300ms by estimating the overhead. If this
> document allows smaller values, it may have to similarly reason about the
> impact.
>
> [CH] I guess it depends on whether the recommendation is to specifically use
> 300ms, or whether the recommendation is to use 300 or lower. RFC 4103 talks
> about specifically using 300ms, so I think we should follow that for T.140 data
> channels. So, I am fine with the text you suggest.
>
> OK, then we may have sorted this out
Good :)
---
> > 5/ Section 5.4 mandates: "Retransmission of already successfully
> transmitted
> > T140blocks MUST be avoided, and missing text markers [T140ad1] SHOULD
> be
> > inserted in the received data stream where loss is detected or suspected." I
> > believe a better wording for the MUST would be "... sucessfully received
> > T140blocks ...", albeit the document does not detail how an
> implementation can
> > indeed fulfill this MUST. Regarding the SHOULD, I assume that "loss
> suspected"
> > could be deterrmined by a heuristic. Could such a heuristic fail and result
> in
> > spurious missing text markers? If so, would a SHOULD be reasonable for
> that?
>
> Regarding the MUST, T.140 does not provide acknowledgement that
> T140blocks have been received. It uses a reliable data channel, so as long as
> the data channel is up and running the sender can only assume that the blocks
> will be successfully transmitted.
>
> Perhaps Gunnar knows more about how receiving implementations would
> "suspect" loss?
>
> The requirement from the T.140 presentation level is that the channel shall
> deliver in order and without duplication.
> Possible loss should be indicated. For RFC 4103, there is a slight risk that
> packets are lost, and if more are lost than
> can be recovered by the redundancy, then a suspected loss has appeared and
> should be indicated in the text presentation.
> There is a chance that something invisible in the text stream was lost. We
> cannot know.
>
> For the t140-usage case, the situation is different. We have reliable delivery in
> order as long as not more than
> about 7 retries are made and nothing is blocking transmission so that the
> watchdog breaks the SCTP associations.
> But that can happen and the draft says that reestablishment shall be tried. At
> that moment it may be hard to know from
> the sender side what was successfully transmitted, because we do not have any
> application level checking delivery. What
> must be avoided is to retransmit something that might have been received. It is
> better to let something be lost and insert
> an indication that something might have been lost.
>
> [MS] I am not really familiar with internals of the SCTP stack. But to me there is
> ambiguity in the term „already successfully transmitted
> T140blocks“. What does that mean? That the T140block has been copied into
> the SCTP send buffer? Or that an IP packet containing this
> data has been transmitted to the peer? I think the former may not immediately
> imply the latter, what has exactly happened may be difficult
> to know. Maybe it would be safer to avoid the word „successfully transmitted“.
> For instance, „Retransmission of already sent T140blocks
> MUST be avoided“ may work around some of the terminology issues.
>
> That is what the text tries to say. It might be a good habit to always insert an
> indicator for suspected loss by the receiver, when the SCTP associations are
> refreshed. I think the current wording allows that and smarter solutions if
> possible. It is good to not require the transmitter to insert the loss marker,
> because that might make it harder to apply security to the transmission chain.
>
> [MS] I was wondering if it is possible that the transmitter erroneously inserts a
> loss marker even if _no_ data was lost, because the receiver wrongly
> „suspected“ loss. As far as I understand, the ITU-T T.140 Addendum specifies
> the loss marker for „discovered data loss“. So, it is not clear to me if using this
> for _suspected_ loss would be consistent with the ITU-T T.140 Addendum
> (which I had to read to actually understand this I-D).
>
> [MS] Anyway, I agree that the current specification allows „smarter“ solutions
> simply by omitting details on (well, granted) corner cases. The question is
> whether the current spec is precise enough to ensure interoperability between
> implementations and to fully describe the characteristics to potential users. I
> don’t understand this use case well enough and therefore I defer that question
> to the TSV ADs.
>
> [CH] Perhaps we could remove "already successfully transmitted T140blocks",
> since there currently is no way to verify that. Maybe something like:
>
> OLD:
>
> "In case of network failure or congestion, T.140 data channels might
> fail and get torn down. If this happens but the session sustains, it
> is RECOMMENDED that implementations tries to reestablish the T.140
> data channels. If reestablishment of the T.140 data channel is
> successful, an implementation MUST evaluate if any T140blocks were
> lost. Retransmission of already successfully transmitted T140blocks
> MUST be avoided, and missing text markers [T140ad1] SHOULD be
> inserted in the received data stream where loss is detected or
> suspected."
>
> NEW:
>
> "In case of network failure or congestion, T.140 data channels might
> fail and get torn down. If this happens but the session sustains, it
> is RECOMMENDED that implementations tries to reestablish the T.140
> data channels. As a T.140 data channel does not provide a mechanism
> for the receiver to identify retransmitted T140blocks, the sender MUST
> NOT retransmit T140blocks unless it has strong reasons to suspect that
> a T140block has been lost. Similarly, as a T.140
> data channel does not provide a mechanism for a receiver to detect
> lost T140blocks, it MUST NOT insert missing text markers [T140ad1]
> unless it has strong reasons to suspect that a T140block has been lost.
> Different mechanisms used by senders and receivers to suspect packet loss
> Is outside the scope of this specification."
>
> That looks better to me.
Good :)
Regards,
Christer
--
last-call mailing list
last-call@xxxxxxxx
https://www.ietf.org/mailman/listinfo/last-call
