Task #5310
closed
Bring ASF ranking implementation into line with technical report
100%
Description
During the course of a separate investigation, we identified a handful of issues with the current ASF face ranking implementation that leads to behavior that deviates from the seemingly expected or intended behavior as described in the technical report.
These include incorrect usage of std::set that leads to faces with identical stats as another preventing the other face from being inserted as well as the incorrect ranking of faces compared to the original design document.
We also aim to resolve issues that lead to unclear or undefined behavior in ranking by introducing the usage of FaceId as a tiebreaker of last resort.
Updated by Davide Pesavento 3 months ago
Can you add a link to the TR for completeness?
Updated by Alexander Lane 3 months ago
Link to technical report is here: https://named-data.net/wp-content/uploads/2016/07/ndn-0042-1-asf.pdf
Updated by Davide Pesavento 3 months ago
- Status changed from In Progress to Code review
- Target version set to 24.next
Updated by Alexander Lane 3 months ago
@Davide I had meant to get in touch with him about it. The lack of a description makes it hard to interpret, I was unclear if it potentially meant the uniqueness enforcement issue I noted or if this was a modification to the probing probability (currently, each face in the order has a decreasing probability based on rank, even if the two of in the same group them have equal SRTT and cost).
Updated by Davide Pesavento 3 months ago
Alexander Lane wrote in #note-5:
I was unclear if it potentially meant the uniqueness enforcement issue I noted or if this was a modification to the probing probability
Sounds like the latter to me. But in that case, is a tiebreaker still needed for the probing rank?
Updated by Alexander Lane 3 months ago
Davide Pesavento wrote in #note-6:
Sounds like the latter to me. But in that case, is a tiebreaker still needed for the probing rank?
It is at least needed in the forwarding case. In the probing case, this still matters for unmeasured faces in terms of reproducibility at least even if we do redesign the probability algorithm significantly.
Updated by Davide Pesavento 3 months ago
- Status changed from Code review to Closed
- % Done changed from 0 to 100