[Anima] Re: spreading the load out in ACP/RFC8994

[Michael Richardson <[e-mail]>]

Toerless Eckert <[e-mail]> wrote:
    > So... how is this supposed to work:


    >             RPL-Root
    >               |
    >              rtr ...  ...  rtr1 rtr2 ...  rtr10
    >               |        |           |
    >            ---+----+---+---+++++---+----- LAN
    >                    |
    >                  rtrX

    > I am on rtrX and wonder, which (subset) of the 10 routers rtr1...rtr10
    > on the LAN to build ACP channels to, right ?

Yes, that's right.

    > Are you telling me i should break confidentiality and export some RPL
    > parameter from inside the ACP into DULL GRASP so that i would know
    > everybody but rtr1 suxx and won't give me a direct connection to the
    > root ?

Yes.  We are doing this in RFC9032.

    > Or else should i randomnly hope that my packets do not have to travel
    > across the LAN because i connect to rtr10, rtr10 itself decided to
    > connect to rtr9, ... and all the way to rtr2 connecting to rtr1 ?

Yes, sure, but with a fan-out of 3, log_3 10 = 2.09590327.
So you'd expect that you need to visit three routers at most.

    > Of course this is a pathological setup, e.g.: i think in actual DC
    > networks withso many paths we just for switches to be ACP enabled as
    > opposed tohaving big, dumb LANs. AFAIK, all the switches support L3,
    > Any of the L2 connectivity sounds more like a service to the hosts, not
    > a limit to the switch hardware.

We need to have a sensible path towards that; so the ACP has to be useful on
"day one".... in the cabinet of systems situation, I see the ACP has being
the ideal network for the server BMCs.

--
Michael Richardson <[e-mail]>, Sandelman Software Works
 -= IPv6 IoT consulting =-                      *I*LIKE*TRAINS*