Title: Aim : Develop a flexible new scheduling methodology which improves fairness by adding knowledge of users
1Wireless Schedulers with Future Sight
viaReal-Time 3D Environment MappingMatthew
Webb, Congzheng Han, Angela Doufexi and Mark Beach
Aim Develop a flexible new scheduling
methodology which improves fairness by adding
knowledge of users future data rates into the
proportional fair scheduling metric.
- Introduction
- New applications, such as Layar and ViewNet
allow augmented reality models to represent the
physical environment in real-time. - ViewNet can produce and store an occupancy grid
associating position to rate, channel state, etc.
and a low-resolution 3D map to permit, e.g.,
coarse RSSI prediction by identifying walls,
doors and windows. - Future data-rates can be estimated by
extrapolating a users recent motion track and
relying on previously stored values of data-rate
at those co-ordinates, or low-resolution
ray-tracing of stored physical structure.
Window marker
Door marker
Occupancy grid
Wall
- Future-Based Scheduling
- In a K-user system, extend user ks proportional
fair (PF) metric to include measures of their
future data-rates
- Scalars ?, ?, ?, ? allow choice of balance
between past, present and future. - Can choose how to define Fk(t) and use in
numerator and/or denominator - Exponentially-weighted decay over N time-slots
into future, similarly to Tk(t) into past.
In numerator denote as 1N
In denominator denote as 1D
- Compute Tk(t) over both past and future windows,
as if user always transmits, for N time-slots.
- Fully compute scheduling at N future times, and
use resulting Tk(t) in PF metric. Effectively, ?
? 0.
Performance
- Future schedulers based on 1N give fairness
improvement over PF for small capacity loss. - Future knowledge in numerator (1N) acts to
smooth out short dips in rate by compensating in
the metric with near-term increases in rate. - Best configuration has future information
weighted less than past (?, ? lt ?, ? ), but does
include both. - Full re-scheduling (3) gives longer-term
average for Tk(t), but statistics of BRAN channel
are stationary. More useful if path-loss is
changing. - 1N 3 makes decisions on the 1N metric, but
long-term average rate is on PF basis, so can
assume wrong users, and capacity falls slightly.
tc tf N 300, 6 users
Simulation parameters 4x4 MIMO-OFDMA with 6 or 10 users, 1024 subcarriers, 768 data subcarriers, guard interval of 176. Transmit power 17 dBm for each user.
Simulation parameters 3000 BRAN C fading realizations with 802.11n path-loss in a 100m-radius cell.
Simulation parameters 48 physical resource blocks (PRBs) of 16 subcarriers are each scheduled separately. Rk(t) is the users mean capacity across the PRB.
- With various system-level parameters, fairness
enhancement for 1N and 1N2 is retained. - General behaviour is familiar from classical PF
scheduler - More users reduces fairness but future-based
schedulers do much better than greedy. - Longer tc and tf trade fairness for capacity.
But 1N 3 loses on both since decisions it
makes are based on more wrong information. - Increasing future horizon, N, also improves
fairness as scheduling metric can take more
future information into account if there is a
near-term dip in rate for a particular user.
? ? 5 ? ? 1
- Conclusions and Future Work
- Future-based schedulers can achieve better
fairness and nearly the same capacity as
classical PF scheduler. - The new scheduling metric including future
knowledge allows a flexible capacity-fairness
tradeoff to be made. - Future-based schedulers with a significant
weighting to the past (?, ?) are the most
successful in this channel model. - Future work Analyse effects of (i) imperfect
future data-rates (ii) motion, i.e. changing
path-loss in channel models.
This work was co-funded by the UK Technology
Strategy Board. We thank all the partners to the
ViewNet project for their help and discussions.