[Bloat] [Ecn-sane] quick question

[Erik Auerswald]

Hi,

On Sat, Aug 26, 2023 at 03:06:09PM +0300, Jonathan Morton via Bloat wrote:
> > On 26 Aug, 2023, at 2:48 pm, Sebastian Moeller via Ecn-sane <ecn-sane at lists.bufferbloat.net> wrote:
> > 
> > percentage of packets marked: 100 * (2346329 / 3259777) = 72%
> > 
> > This seems like too high a marking rate to me. I would naively expect
> > that a flow on getting a mark scale back by its cwin by 20-50% and
> > then slowly increaer it again, so I expect the actual marking rate
> > to be considerably below 50% per flow...
> 
> > My gut feeling is that these steam flows do not obey RFC3168 ECN
> > (or something wipes the CE marks my router sends upstream along the
> > path)... but without a good model what marking rate I should expect
> > this is very hand-wavy, so if anybody could help me out with an easy
> > derivation of the expected average marking rate I would be grateful.
> 
> Yeah, that's definitely too much marking.  We've actually seen this
> behaviour from Steam servers before, but they had fixed it at some
> point.  Perhaps they've unfixed it again.
> 
> My best guess is that they're running an old version of BBR with ECN
> negotiation left on.  BBRv1, at least, completely ignores ECE responses.
> Fortunately BBR itself does a good job of congestion control in the
> FQ environment which Cake provides, as you can tell by the fact that
> the queues never get full enough to trigger heavy dropping.
> 
> The CUBIC RFC offers an answer to your question:
> [small screenshot attached to email]

I find the attached screenshot quite unreadable.  It seems to be
taken starting from the paragraph above section 5.2 of RFC 9438
<https://www.rfc-editor.org/rfc/rfc9438#section-5.2>.  In UTF-8 text it
looks as follows:

------------------------------------------------------------------------

   _C_ determines the aggressiveness of CUBIC in competing with other
   congestion control algorithms for bandwidth.  CUBIC is more friendly
   to Reno TCP if the value of _C_ is lower.  However, it is NOT
   RECOMMENDED to set _C_ to a very low value like 0.04, since CUBIC
   with a low _C_ cannot efficiently use the bandwidth in fast and long-
   distance networks.  Based on these observations and extensive
   deployment experience, _C_=0.4 seems to provide a good balance
   between Reno-friendliness and aggressiveness of window increase.
   Therefore, _C_ SHOULD be set to 0.4.  With _C_ set to 0.4, Figure 7
   is reduced to

                                            4 ┌────┐
                                            ╲ │   3
                                             ╲│RTT
                       AVG_W      = 1.054 * ────────
                            cubic            4 ┌──┐
                                             ╲ │ 3
                                              ╲│p

                                  Figure 8

   Figure 8 is then used in the next subsection to show the scalability
   of CUBIC.

5.2.  Using Spare Capacity

   CUBIC uses a more aggressive window increase function than Reno for
   fast and long-distance networks.

   Table 3 shows that to achieve the 10 Gbps rate, Reno TCP requires a
   packet loss rate of 2.0e-10, while CUBIC TCP requires a packet loss
   rate of 2.9e-8.

      +===================+===========+=========+=========+=========+
      | Throughput (Mbps) | Average W |  Reno P | HSTCP P | CUBIC P |
      +===================+===========+=========+=========+=========+
      |                 1 |       8.3 |  2.0e-2 |  2.0e-2 |  2.0e-2 |
      +-------------------+-----------+---------+---------+---------+
      |                10 |      83.3 |  2.0e-4 |  3.9e-4 |  2.9e-4 |
      +-------------------+-----------+---------+---------+---------+
      |               100 |     833.3 |  2.0e-6 |  2.5e-5 |  1.4e-5 |
      +-------------------+-----------+---------+---------+---------+
      |              1000 |    8333.3 |  2.0e-8 |  1.5e-6 |  6.3e-7 |
      +-------------------+-----------+---------+---------+---------+
      |             10000 |   83333.3 | 2.0e-10 |  1.0e-7 |  2.9e-8 |
      +-------------------+-----------+---------+---------+---------+

        Table 3: Required Packet Loss Rate for Reno TCP, HSTCP, and
                   CUBIC to Achieve a Certain Throughput

   Table 3 describes the required packet loss rate for Reno TCP, HSTCP,
   and CUBIC to achieve a certain throughput, with 1500-byte packets and
   an _RTT_ of 0.1 seconds.

------------------------------------------------------------------------

(extracted using:

    wget -q -O- 'https://www.rfc-editor.org/rfc/rfc9438.txt' \
    | sed -En '/^ +_C_ determines the aggressiveness of CUBIC/,/of 0\.1 seconds\.$/p'

)

> Reading the table, for RTT of 100ms and throughput 100Mbps in a single
> flow, a "loss rate" (equivalent to a marking rate) of about 1 per
> 7000 packets is required.  The formula can be rearranged to find a
> more general answer.

HTH,
Erik