Support weights_determined_by_switch field in P4Info

[matthewtlam]

The switch has its own hash algorithm that already exists (ex. CRC, Identity, etc.). Traditionally, these are not defined in the P4Info in any way since it doesn’t affect the program.

We wish to introduce this new field ,weights_determined_by_switch, to the P4Info to denote to the control plane that programming will have no effects on the weights. We wish to leverage the switch using its own hash algorithm to determine the weights.

cc: @jonathan-dilorenzo, @smolkaj for vis

[chrispsommers]

Please provide a summary description, thanks.

[matthewtlam]

Other design considerations that we had include:

• Expose a new type of hash algorithm in V1Model (needed for our purpose) and PSA (for standard compliance). The P4Info knob weights_determined_by_switch would then be set based on this hash algorithm.

Other questions

• Do we care about some groups having a particular weight defined by the switch and then others groups that the switch can program?

EDIT: changed the link to reference P4C version

cc: @chrispsommers, @jafingerhut, @antoninbas

[antoninbas]

@matthewtlam you shared an internal Google link

You could introduce a standard P4Runtime annotation to control whether that boolean should be set in the P4Info (similar to @max_group_size)

[jonathan-dilorenzo]

Thanks @antoninbas! That would be our suggestion if we want to do this boolean approach. The main question, I think, would be if there is something more expressive we might want to put here instead, though I can't immediately think of other ways in which the hash algorithm might affect what the controller programs... I suppose a 'per-group' hash algorithm (including "You choose weights") would be a way to go (as we sort of discussed on Angela's PR), but I don't actually know of any switch with close to that expressive power ;).

[smolkaj]

"switch using its own hash algorithm to determine the weights"
I'm a bit confused by this part. Is it the job of the hash algorithm to determine the weights? Presumably it is the job of some sort of load balancing protocol/algorithm? And the hash algorithm is orthogonal?

[chrispsommers]

"If weights_determined_by_switch, seems like an incomplete phrase."

How about "If weights_determined_by_switch is true"