[e2e] Ressource fairness of PF scheduling and similar schemes in the presence of asymmetric fading.

[Detlef Bosau]

Once again me :-)

At the moment, I see two areas of interest in opportunistic scheduling 
as it is used in HSDPA systems. (At least, they are of interest for me. 
I don´t know the opinion of others people.)

For quite some time, proportional fair scheduling was state of the art. 
Recently, some alternatives have been proposed.
The main shortcoming in PF scheduling is its behaviour in the presence 
of asymmetric fading.
Refer to
http://perso.rd.francetelecom.fr/bonald/Pub/SB.pdf.

Thomas Bonald, "A Score-Based Opportunistic Schedulerfor Fading Radio 
Channels".

Intuitively, the reason for the problem  is that SNR values for 
different UEs may vary on different average values and with different 
variation. Hence to compare SNR values for different UEs is often 
comparing peaches and oranges.

To my knowledge, we have to approaches to overcome this problem.

One was developed within the Eurane Project, *

Fair Channel Quality-Based Scheduling Scheme for HSDPA System*
Al-Manthari, B.; Nasser, N.; Hassanein, H.
Computer Systems and Applications, 2006. IEEE International Conference on.
Volume , Issue , March 8, 2006 Page(s): 221 - 227
Digital Object Identifier  

The idea is to use statistices "(SNR value - average) / standard 
devation" instead of the original SNR values.
So, for any UE the considered statistics share a common expectation (0) 
and a common variation (1) and are hence comparable.

Another approach is the aforementioned one by Thomas Bonald, who 
basically (please correct me, if I have a misconception here) uses 
quantiles / ranks as statistics to compare SNR values for different UEs.

In fact, quantiles are not mentioned explicitely in Bonald´s work, but 
practically, this is the idea behind his approach: The actual SNR is 
ranked within a set of n most recent SNR values.

Im not sure yet, whether we achieve comparable statistics here.

The other issue is ressource fairness / QoS in opportunistic scheduling. 
I did some experiements with a token based approach during the last 
weeks - however, the problem is to achieve stability because of the 
uncomparable SNR statistics.
If we had SNR statistics which are i.i.d., a token based approach could 
be helpful here, but I´m afraid that´s hardly to achieve.
So, at the moment, I plan to investigate some rate based approach.

The idea is to have some kind of "modified round roubin" scheduling, which
- enforces fair (or if QoS applies: according to contract) ressource 
usage for any flow in the long run
- allows to use higher or lower ressource usage in the short run in 
order to adapt the system to favourable or poor channel conditions.

I´m curious whether there are other people interested in this matter, 
especially people from Germany.

Detlef

-- 
Detlef Bosau                          Mail:  detlef.bosau at web.de
Galileistrasse 30                     Web:   http://www.detlef-bosau.de
70565 Stuttgart                       Skype: detlef.bosau
Mobile: +49 172 681 9937