On Wed, Apr 12, 2017 at 10:33 AM, Toke Høiland-Jørgensen <toke@toke.dk> wrote:
> Matias Richart <mrichart@fing.edu.uy> writes:
>
>> Hi to all! I occasionally follow the wifi-fast list and I have just
>> found this initiative.
>> I've been working with ns3 rate control for several years and I would
>> also like to contribute if there is an ns3 approach.
>
> Awesome! Welcome :)

Hi Pasquale and Matias, it's great to see some new names on the list! :)

>>> And is there a concept "retry chains" for the rate selection
>>> algorithm (where a sequence of rates to try are picked at once when a
>>> decision is made)?
>>
>> It exists an implementation of Minstrel and Minstrel HT, which uses
>> the concept of retry chains, but this is implemented in the same
>> algorithm, not as part of the MAC layer.
>>
>> In summary, we can simulate the retry chain behavior. Currently, there
>> are functional implementations of Minstrel and Minstrel HT which I
>> think work well.
>
> Right, excellent. The retry chain and the inability to re-calculate all
> probabilities for every packet are some pretty hard constraints on real
> hardware, so having simulation work in a similar way is most likely
> quite central for carrying over the simulation results to a Linux
> implementation.
>
>>>I think the two main things we are trying to figure out are the
>>>correlations between different rates. Which involves answering
>>>questions like:
>>>
>>>1. What correlations exist between the success probabilities of sending
>>>   at different rates. I.e., can we always assume that if a
>>> transmission
>>>   fails at a low rate (more robust encoding), it would also have
>>>   failed at a higher rate (or conversely, if it succeeds at a high
>>>   rate, it would also have succeeded at the lower rate). Does this
>>> hold
>>>   within the same MIMO configuration? What about between different
>>> MIMO
>>>   configurations?
>>
>> In my opinion, this is quite easy to implement. I'm thinking on an
>> experiment with an static deployment and testing all possible rates.
>
> Yes, that was my thought as well, and I figure this is easier to do in
> simulation.

One thought: Since we're doing simulation, would it be possible to
compute the packet success probability of every rate, for every
packet? I mean so that we get to know the "counterfactual": "Our
algorithm chose MCS-13, which according to simulation had a success
probability of 0.5 for this specific transmission, but according to
the simulation MCS-15 had a success probability of 0.7 *for this same
transmission*"? Then we could calculate correlations and regret and
similar very easily...

/Björn