Re: [6tisch] MSF traffic adaptation under stress

[Tengfei Chang <[e-mail]>]

Hi David,

I replied inline:

On Fri, Apr 10, 2020 at 12:27 AM David Hauweele <[e-mail]>
wrote:

> Dear 6TiSCH,
>
> In the last few months, we performed a small scale study of the
> behavior of 6TiSCH's minimal scheduling function (MSF) under stress. We
> were especially interested in the dynamics of MSF's automated
> adaptation to traffic load. Some of our conclusions have been
> summarized in a paper that we submitted to IEEE ISCC 2020. A copy of
> the evaluation section of our submission is attached to this mail.
>

Congrats on the published paper!

>
> Among our surprising findings, we observed that
>
> 1) The time required for MSF to adapt to a traffic change depends on
> the current number of cells allocated. To give an example, it takes
> much longer to go from 0 to 10 cells than from 10 to 20 cells. We
> attribute this behavior to the way MSF measures cell occupancy by
> counting the number of used and passed cells. Adaptation only occurs
> when the number of passed cells reaches a maximum. However, with light
> cell allocation, it takes longer to reach that maximum than with higher
> cell allocation.
>

Cool! I have the similar results when evaluate the MSF performance.
There are two things influencing the response time to the traffic changes
1. The value of MAX_NUM_CELLS, which you mentioned. It is configurable. By
setting it to a small value, the  response time can be reduced in case the
traffic load changes frequently.
2.  The number of cells to be added/deleted each time.  In MSF, for the
simplicity, we only add/delete 1 cell every time. An advanced version of
MSF could add/delete cells according the percentage of cell usage.

>
> 2) MSF can lead to severe over-provisioning, which can be harmful in an
> environment where the resources are scarce. We noticed that releasing
> cells was especially hard for MSF due to the fixed hysteresis
> thresholds. Indeed, the estimated cell occupancy must drop below 25%
> for cells to be released.
>

The over-provisioning is designed intentionally to avoid the fluctuation of
6P transactions.
It costs additional 50% cells in average, as a trade-off, it could reduce
the cost of sending 6P frames, and the latency caused by 6P transaction.
No sure if there is a perfect way to cover every aspects?

>
> We already have ideas to improve MSF's traffic adaptation mechanism,
> that we plan to put under test in the coming weeks. We can also provide
> you with more details if you wish. If you see interest in our evaluation
> and proposals, we are eager to discuss this further with the WG.
>

Very interesting to see your approach to improve MSF !
Referring to the MSF standardization process , as said by Pascal, we won't
be able to made big changes.
We can adapt some minor changes.
Unless there is a flaw in the draft, I would prefer to keep the draft as it.

Tengfei


>
> Best regards,
> David, Bruno, Aris and Georgios
> _______________________________________________
> 6tisch mailing list
> [e-mail]
> https://www.ietf.org/mailman/listinfo/6tisch
>


-- 
——————————————————————————————————————
Stay healthy, stay optimistic!

Dr. Tengfei, Chang
Postdoctoral Research Engineer, Inria

www.tchang.org/
——————————————————————————————————————