Re: [Idr] Choice of Large vs. Extended Community for Route Leaks Solution

[Brian Dickson <[e-mail]>]

On Thu, Apr 1, 2021 at 2:06 PM Sriram, Kotikalapudi (Fed) <
[e-mail]> wrote:

> There may be a knob that AS operators have for permitting transitivity,
> but we need to look at measurements to understand whether or not operators
> actually allow transitivity to EC and LC.
>
> NIST BGP measurements (thanks to my colleague Lilia Hannachi) were shared
> on the GROW list in May 2020:
> https://mailarchive.ietf.org/arch/msg/grow/JPD1-hhSvVXIZbUlNQ_1hmzD6IA/
>
> A portion is copied below. The AS path length (# unique ASes)
> distributions for BGP updates with Communities (Regular, Large, and
> Extended) are shown here. It is evident that both LC and EC propagate
> multiple AS hops. Mass stripping of LC or EC at the first hop is not
> evident.  The peak happens at AS path length 4 or 5 and that is good. That
> is the behavior that is helpful for route leak solution. The solution can
> still function even if some ASes strip. We can do some more detailed
> studies if needed.
>
> *********************************************************************
> RIPE-RIS: Community ANALYSIS (Collector : rrc03 From 2020-04-30 00:00 To
> 2020-04-30 00:55)
> *********************************************************************
> # Updates = 1075583 (Total)
> # (Regular) COMMUNITY = 859239 (79.89%)
> AS path length distribution =    1: 170 (0.02%)    2: 44803 (5.21%)    3:
> 141072 (16.42%)    4: 276271 (32.15%)    5: 238325 (27.74%)    6: 114158
> (13.29%)    7: 31365 (3.65%)    8: 9018 (1.05%)    9: 2690 (0.31%)    10:
> 811 (0.09%)    11: 358 (0.04%)    12: 169 (0.02%)    13: 22 (0%)    14: 7
> (0%)
>
> # LARGE_COMMUNITY = 152818 (14.21%)
> AS path length distribution =    2: 5655 (3.7%)    3: 17205 (11.26%)    4:
> 54372 (35.58%)    5: 45492 (29.77%)    6: 22065 (14.44%)    7: 6422 (4.2%)
>   8: 1068 (0.7%)    9: 397 (0.26%)    10: 71 (0.05%)    11: 35 (0.02%)
> 12: 26 (0.02%)    13: 6 (0%)    14: 4 (0%)
>
> # EXTENDED COMMUNITIES = 44606 (4.15%)
> AS path length distribution =    2: 2269 (5.09%)    3: 7435 (16.67%)    4:
> 17657 (39.58%)    5: 11600 (26.01%)    6: 3967 (8.89%)    7: 1221 (2.74%)
>   8: 371 (0.83%)    9: 57 (0.13%)    10: 19 (0.04%)    11: 8 (0.02%)    12:
> 1 (0%)    13: 1 (0%)
> *********************************************************************
>

One main issue will be the use of LC vs WC vs regular communities, among
the largest networks (e.g. networks nominally at the tier-1 designation).
Going from e.g. a wikipedia page with list of ASNs, that would be all or
most of these ASNs:
7018, 1299, 1239, 3320, 6453, 6461, 3356, 3549, 2914, 3257, 3491, 701,
6762, 3491.

Can you get a top-level summary of LCs with GA values from that list, and
whatever the corresponding thing would be for EC or WC?
E.g. How many unique LC values are there that match 3320:* ? Maybe a table
with ASN vs common LC count would help illustrate this.

If the majority (or vast majority) are already using LCs, and many fewer
are using WC/EC, this definitely should tilt the argument towards LCs.

Any of these networks which is using LCs extensively will very likely to be
unwilling to use anything other than a WKLC (and this use of LCs was to a
significant degree the result of requests from operators that use LCs to do
it with LCs.)

Brian