Oscillations in over-utilization and under-utilization cycle can be added under the evaluation criteria

[altanai]

Since this is a problematic behaviour because it leads to an inefficient use of network resources and can cause variable/unreliable performance ( high speed followed by low speed so on ). Maybe new Cc algorithms can focus on stability in this area.

[gorryfair]

I agree that oscillations in the rate can really impact performance, and that this might be useful as a performance result. Do you think this translates in to harm for other flows (e.g., starvation) or poor use of the available capacity?

[altanai]

Exactly. I think such cycles would be detrimental to end-user experience, any telemetric system that a network service provider may be using as well as the both the points you mentioned.

[huitema]

We have to be a bit careful about recommending stability. I remember Van Jacobson explaining that yes, a more stable version of Reno would be a bit more efficient, but then it would also not adapt so quickly to changes in network conditions. There might be a tradeoff there.

So maybe we need to be very clear about the characteristics of the oscillation that we do not want. @altanai are you thinking about something like the "saw tooth" patterns of RENO, or the similar patterns in Cubic? Or do you have some example of worse oscillations?

[martinduke]

I'm not sure we need a new criterion here. Under-utilization is not a problem for network stability. Over-utilization is well-covered by our bufferbloat metric.

Even an algorithm that sawtooths around the correct link capacity will experience more buffering than a smoother algorithm.