Re: [PATCH ha-manager 7/7] manager: auto rebalance: only balance under actual node resource pressure

[Michael Ryom <Michael@RyomHerold.dk>]

Hi Daniel,

thanks for the detailed explanation - I agree with your points and
have dropped this patch from v2.

In particular, that a node-local PSI value can say "something is
wrong here" but not what should go where, is what I ran into myself
when trying to couple the candidate selection to the pressure: without
a model of what each guest needs and gets, the gate only decides
*when* to run a balancer that still optimizes for an even load.

For what it's worth, the measurements from the test cluster pointed in
the same direction: over 23 hours, the gate opened 45 times, but only 5
of these openings were backed by a one-minute average above the
threshold - the avg10 value sampled once per round mostly caught short
spikes, so a gate would have needed a sustain requirement on top.

I'll come back with a separate proposal (RFC) for the "what and where"
part, based on per-guest measurements (CPU time a guest wanted but did
not get, from the schedstat run_delay of its threads) rather than on
node-local pressure, and with memory treated as a constraint instead
of something to balance. That is independent of this series.

Best regards,
Michael