Re: [tcpm] More motivating scenarios for tcpm-ack-pull

["Carles Gomez Montenegro" <[e-mail]>]

Hi Bob,

Once again, thanks a lot for all the very useful feedback!

Please find below some inline responses.

> Carles, Jon, (re-sending, this time without accidentally omitting tcpm,
> also see couple of addenda tagged [Bob adds:])
>
> I'm generally supportive of a mechanism to suppress delayed ACKs from
> the receiver (in any transport protocol including TCP). So thank you.
>
> You might want to borrow at least some of the text for additional
> motivating scenarios for such a facility from the first two posts on a
> thread for a similar facility in QuiC:
> https://github.com/quicwg/base-drafts/issues/1978
>
> Also pasted at the end.
>
> Some is not relevant to your specific ack-pull scheme, 'cos it's
> proposing a more general facility in QUIC for the sender to the receiver
> alters its ACK ratio. (theoretically there are more bits available in
> QUIC, but this particular request has been put on hold while someone
> goes and finds them ;)

Thanks a lot. This is indeed very useful!

As you explain, the more general facility for QUIC allows to suppress
Delayed ACKs and also allows to modify the ACK ratio in a more general
way.

Regarding the Delayed ACK suppression discussion, the additional
motivation is definitely important.

>       Another motivating example (rather niche)
>
> Coincidentally, Just this morning I published a tech report that had to
> use the hack of overlapping a byte from the previous segment to force a
> quick ack:
> TCP Prague Fall-back on Detection of a Classic ECN AQM
> <https://arxiv.org/pdf/1911.00710.pdf#page9> (the link takes you to Fig
> 2, which illustrates the hack)
> I'd rather not have to do these sorts of hacks - it really messes with
> the segmentation code.

This is very useful feedback as well, and a nice coincidence!

Regarding the hack, sending an old byte may be even worse when there is
not a next data segment readily available for sending, which would lead to
sending a whole packet just to carry an old byte.

>       Middlebox traversal of bit 6
>
> I recall that someone did a study of traversal of each of the reserved
> flags (Marcelo Bagnulo maybe?). I don't think it was that good [Bob
> adds: I mean traversal - not the study!].
> A good search engine might find it ;)

I wasn't able to find the study, but this is an important consideration.

One might think that if new functionality is defined and standardized,
related middlebox traversal of it should improve over time. Perhaps that
might be optimistic...

>       More alternative approaches:
>
> 1/ At one stage, we tried to include a bit for Delayed Ack control in
> another protocol called Accurate ECN. We were persuaded to take it out,
> because it was mission creep (outside the scope of the protocol we were
> working on):
> https://tools.ietf.org/html/draft-kuehlewind-tcpm-accurate-ecn-03#appendix-B.4

I was unaware of this, thanks!

This is actually very aligned with what we are proposing...

> 2/ Have you considered using the Urgent Pointer in a novel way? I.e. a
> non-zero Urgent Pointer field, even tho the URG flag is zero.
> [Bob adds:] From memory, traversal was pretty good - much better than
> bits 4-6. But I think Windows treated it as a potential attack. I
> believe Fernando Gont did a study on how 'invalid' TCP header fields are
> handled.
>
> You could assign, say, 3 bits of the urgent pointer for a variable
> ack_exp that is log base 2 of the ack ratio the sender would like. Then
> the sender can request the receiver uses ack_ratio` = 2^ack_exp, as in
> the QUIC proposal below,
> With ack_exp = 0, you get ack_ratio = 2^0 = 1, which has the same effect
> as ack-pull. But you also have a more general facility for high speed
> machines to widen out the ack ratio.
>
> If this works, you could open up a registry for the other 13 bits. So
> instead of using up a precious bit, you generate 13 more ;)

Nice! :)

Actually, we had not considered such an idea.

In your proposal above, if a segment has the URG flag set to 1, would then
the 3 ack_exp bits still be used as ack_exp bits, thus reducing the size
of the urgent pointer in practice to 13 bits?

> See:
> https://tools.ietf.org/agenda/90/slides/slides-90-tcpm-10.pdf#page=5
>
> I think there were problems with some implementations and middleboxes
> (bound to be). I've checked the minutes of the IETF meeting where that
> slide was presented
> <https://www.ietf.org/proceedings/90/minutes/minutes-90-tcpm>, but
> there's no push-back in there. I suggest you delve back into the
> archives of the tcpm mailing list around the time of that meeting to
> find out if there were any show-stoppers.

Thanks for the pointers!

Sure, we'll check the tcpm mailing list archive.

Cheers,

Carles


> Cheers
>
>
>
> Bob
>
>
> Paste from QUIC thread:
> /\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\/\
>
>
>
>       Problem/goals:
>
>  1. Getting up to speed fast without inducing much queue is one of the
>     main areas where latency reductions are needed (for short and long
>     flows). The sender needs frequent ACKing during this phase in many
>     approaches (hystart, etc). A Linux TCP receiver starts with
>     ack_ratio 1 and uses heuristics to identify the end of the sender's
>     slow-start (or a re-start after idle). This has ossified a certain
>     slow-start behaviour into all TCP senders (whether Linux or not). We
>     are trying to new sender behaviours (paced chirping being an
>     example), but we need a way for the sender to suppress delayed ACKs.
>  2. At the other end of the scale, for long-running flows, many current
>     CC approaches (e.g. BBR) use pacing not ACK-clocking so they can use
>     very few ACKs per RTT. And fewer means less cache ejections for GRO.
>     But the receiver doesn't know what CC the sender is using, so it
>     doesn't know how many or few is too many or too few.
>  3. Some middleboxes and the majority of link technologies (e.g. DOCSIS,
>     LTE, Satellite) thin TCP ACKs when they detect the upstream is
>     filled with TCP ACK stream(s), which are unresponsive. Otherwise the
>     ACKs constrain downstream throughput (and any upstream data either
>     in the flow itself, or in others). We don't want these links to
>     attempt to guess which are the QUIC ACKs and try to thin them. QUIC
>     can and should do this itself. QUIC has all the machinery for the
>     sender CC to detect ACK congestion, but not the protocol to tell the
>     receiver to do the thinning. This protocol addition would provide a
>     sufficient hook for hosts to unilaterally add this to their CC
>     behaviour.
>
>
>       Some scenarios where the sender's preferred ack_ratio changes
>       through the connection:
>
>  1.
>
>     A sender CC that wants the receiver to turn off DelAcks during
>     flow-start (e.g. it's using hybrid slow-start as in Cubic and wants
>     to get delay measurements more frequently) sets ack_exp=0 during
>     flow-start (ack_ratio=1), then increases ack_exp during congestion
>     avoidance. If it goes idle, then re-starts, it would set ack_exp=0
>     again.
>
>     Note on heuristics: A Linux TCP receiver currently uses a heuristic
>     to determine when the sender has exited slow-start. However,
>     heuristics -> ossification. A Linux receiver's heuristic only works
>     with the current pattern of slow-start. In TCP, when we tried to
>     improve the pattern on the sender (paced chirping), the heuristic on
>     the receiver killed us.
>
>  2.
>
>     Imagine a paced sender has hardware generic receive offload (GRO),
>     so for a long-running flow it doesn't want a high rate of QUIC ACKs
>     that are opaque to GRO. Let's say it would prefer at least 8 ACKs
>     per RTT. Again, it starts with ack_exp=0, but in congestion
>     avoidance it would use:
>
>     ack_ratio <= cwnd_in_packets/8
>
> Upshot: By remote controlling the receiver, the server offloads nearly
> all the ACKs from large downloads, but still focuses its ACK-receiving
> resources on getting each client up to speed.
>
> Note that calculation of ack_ratio lends itself to fast integer
> arithmetic.
>
>
>
>
>
>
>
> Bob
>
>
> --
> ________________________________________________________________
> Bob Briscoehttp://bobbriscoe.net/
>
>