Queues: generalize PIFOs

[anshumanmohan]

At present our PIFOs can only handle:

• Two flows.
• A policy that attempts to give those two flows equal shares.

These limitations of PIFOs also limit PIFO trees.

There is room for generalization on a few axes. I see these as orthogonal, though of course they'd be more powerful if combined. I think these are doable even just using our approach where a PIFO merely orchestrates some n FIFOs, where n can be determined in advance.

• [ ] A PIFO that handles n flows and affects the policy du jour on them.
• [ ] A PIFO that allows strict prioritization among its children: strictly prefer flow A over flow B over flow C, and so on. This PIFO would still be work-conserving: if flow A were silent and flow B had items, we'd send B's items until A became chatty again, and we'd send C's items only if A and B were both silent. This generalizes easily to n flows.

[sampsyo]

This all seems quite sensible! Except that I don't quite understand one thing: what does "strict prioritization" mean? Is it an alternative to fairness?

[anshumanmohan]

Ah, sorry! I'll edit the text above to explain, but yes, this is a different scheduling policy that I think is easily within our grasp using just our simple approach. The idea is to strictly prefer flow A over flow B over flow C, and so on.