1. Distributed Channel Assignment for minimal islands of connectivity
Month Year
doc.: IEEE yy/xxxxr0
Mar 2009
Distributed Channel Assignment for minimal islands of connectivity
Date:
Authors:
Raghuram Rangarajan, Cisco Systems
John Doe, Some Company

2. Mar 2009
The proposal for TGaa to limit the addressed OBSS scenarios is unsatisfactory
It has been suggested that TGaa not address all OBSS scenarios
ie not address larger “extents”
However, user expectations mean that this is only a useful strategy if it works “most of the time”
Assumption: users will not accept their video working perfectly for 29 days and poorly on day 30 (invariably the day of a big event)
The open question is whether OBSS scenarios outside the scope of current proposals occur “frequently” enough to matter
The unfortunate answer is they do in realistic scenarios and so current proposals to address OBSS are unsatisfactory
This is before even considering some of the other “hard problems related to OBSS
Raghuram Rangarajan, Cisco Systems

3. Mar 2009
Simulation shows scenarios outside scope of current proposals occur frequently enough
An AP layout in a typical residential location was simulated using:
A set of reasonable model parameters
The “extent” of “connectivity graphs”
A variety of channel allocation algorithms
Previous results show large extents result from the use of a “classic” channel selection algorithm in a typical scenario
New simulation work with refined channel selection algorithms shown large extents in many typical (and realistic) scenarios
Raghuram Rangarajan, Cisco Systems

4. An AP layout in a typical residential location was simulated
Month Year
doc.: IEEE yy/xxxxr0
Mar 2009
An AP layout in a typical residential location was simulated
Pool
Distance (m)
Raghuram Rangarajan, Cisco Systems
John Doe, Some Company

5. A set of reasonable model parameters was selected
Mar 2009
A set of reasonable model parameters was selected
AP every 700 sq. ft. (approximate apartment size)
Or AP every 1400 sq. ft. (for larger apartments)
Transmit power 15dB
Transmit & receive gain 2dB
General loss at 1m is 47dB
Path loss 35dB every 10m
Floor decay 10dB per floor
5-7dB shadowing variation
APs turn on randomly
Raghuram Rangarajan, Cisco Systems

6. The simulation uses the “extent” of “connectivity graphs”
Mar 2009
The simulation uses the “extent” of “connectivity graphs”
Define a “connectivity graph” as:
an `island' of APs where every AP can be connected via edges to every other AP
Define “extent” of a connectivity graph as:
the minimum (across APs in the connectivity graph)
of the maximum (across pairs of APs)
of the shortest path (between a given pair of APs)
Raghuram Rangarajan, Cisco Systems

7. The simulation uses the “extent” of “connectivity graphs”
Mar 2009
The simulation uses the “extent” of “connectivity graphs”
Graph with 11 nodes.
Extent of the graph is 3
Graph with 8 nodes.
Extent of the graph is 1
Raghuram Rangarajan, Cisco Systems

8. The simulation is based on a variety of channel allocation algorithms
Mar 2009
The simulation is based on a variety of channel allocation algorithms
There are a wide variety of channel allocation algorithms to establish a minimal extent network
The simulation explores a variety of obvious and optimal algorithms based on a basic framework
The algorithm framework is as follows:
The AP scans all channels and records the number and RSSIs of all APs per channel
The AP may also request the connectivity graph and related parameters of other APs
e.g. via Public Action request/response frames
The AP selects the best channel according to a first, second and third sort key
Raghuram Rangarajan, Cisco Systems

9. The simulation is based on a variety of channel allocation algorithms
Mar 2009
The simulation is based on a variety of channel allocation algorithms
Typical 1st sort keys
Typical 2nd sort keys
Typical 3rd sort keys
1a) Channel with 0 APs
1b) Channel with minimum connectivity extent (after assigning the AP to that channel)
1c) Channel with minimum number of APs
2a) Channel with minimum number of edges in the connectivity graph (after assigning the AP to that channel)
2b) Channel with minimum number of vertices in the connectivity graph (after assigning the AP to that channel)
2c) Null (i.e. skip to the third sort key)
3a) Channel with minimum maximum (across neighboring APs) RSSI (i.e. join the most distant connectivity graph)
3b) Channel with maximum maximum (across neighboring APs) RSSI (i.e. join the nearest connectivity graph)
3c) Random
Raghuram Rangarajan, Cisco Systems

10. The simulation is based on a variety of channel allocation algorithms
Mar 2009
The simulation is based on a variety of channel allocation algorithms
The connectivity metrics are calculated after adding the AP to that channel
Since multiple small disjoint graphs may be connected into one large graph by adding the AP.
Different algorithms are created by selecting different sort keys.
1a-2c-3a is a classic algorithm that tries to find an empty channel, or the channel with most distant APs.
1b-2c-3b is better to minimize the connectivity extent
Cost and complexity of obtaining metrics is assumed to be zero
Yet clearly obtaining up-to-date topology information is complicated, incurs overhead, and any information may be readily spoofed
Most complicated after a power cut, where APs are dynamically choosing channels, changing the topology moment by moment
Raghuram Rangarajan, Cisco Systems

11. The simulation is based on a variety of channel allocation algorithms
Mar 2009
The simulation is based on a variety of channel allocation algorithms
In the following graphs, we compare the performance (in terms of connectivity extents) of the following algorithms
Classical approach: 1a-2c-3a
One example of extent-optimal algorithm: 1b-3b
1b-3c
1b-3a
1b-2b-3b
1b-2b-3a
Other variations could also be tried and evaluated for performance
Raghuram Rangarajan, Cisco Systems

12. Mar 2009
Previous work shows large extents result from the use of classic 1a-2c-3a algorithm in a typical scenario
The use of the classic 1a-2c-3a algorithm results in a large extent
See 08/1250
Centralized scheduling in this case is very challenging
Can we do better?
Raghuram Rangarajan, Cisco Systems

13. Mar 2009
New simulation work with new algorithms show large extents in many typical scenarios
The results of the simulations show large extents in many typical scenarios
All algorithms simulated at 2.4GHz (with three channels) result in unsatisfactorily large extents
Insensitive APs using 20MHz channels worked most of the time at 5GHz, but it is not a realistic scenario
Insensitive APs using 40MHz had an extent >1 up to 8% of the time at 5GHz, but it is not a realistic scenario
Sensitive APs using 40MHz commonly had large extents at 5GHz, with associated challenges
This is even before addressing denser situations like Manhattan or Hong Kong
Raghuram Rangarajan, Cisco Systems

14. Mar 2009
All algorithms simulated at 2.4GHz (with three channels) result in unsatisfactorily large extents
Assumptions
Channels =
AP sensitivity = -82dBm
All algorithms simulated at 2.4GHz (with three channels) result in large extents
This is unsatisfactory in the context of OBSS resolution
Raghuram Rangarajan, Cisco Systems

15. Mar 2009
Insensitive APs using 20MHz channels worked most of the time at 5GHz, but it is not a realistic scenario
Assumptions
Channels = 20,
Sensitivity of APs = -82dBm (insensitive)
All algorithms perform very well:
Optimal algorithm: 100% extent of 1
Classic algorithm: 2% with extent >1, ie 2% are problematic
Unfortunately 11n at 20 MHz is less realistic
Only classic and optimal plotted as all minimal extent approaches are similar
Raghuram Rangarajan, Cisco Systems

16. Mar 2009
Insensitive APs using 40MHz had an extent >1 up to 8% of the time at 5GHz, but it is not a realistic scenario
Assumptions
Channels = 10,
Sensitivity of APs = |-82dBm (insensitive)
The classic algorithm performs poorly
Other algorithms perform well but still leave 4-8% beyond an extent of 1, with associated scheduling and sync challenges
However, in the era of 11n, -82 dBm sensitivity is less realistic
Raghuram Rangarajan, Cisco Systems

17. Mar 2009
Sensitive APs using 40MHz commonly had large extents at 5GHz, with associated challenges
Assumptions
Channels = 10,
Sensitivity of APs = |-95dBm (sensitive)
All algorithms performs poorly, with extents of 2-3-4
Sync and scheduling are very challenging in this environment
Raghuram Rangarajan, Cisco Systems

18. Mar 2009
The current OBSS proposals are unsatisfactory given they do handle large extents
Islands of connectivity with large extent happen in a sufficiently interesting number of cases that we need to be concerned about them
And yet the current OBSS proposals do not address them, except by suggesting dropping back to EDCA
However, simply dropping to EDCA remains problematic
Either 11aa’s OBSS work adds nothing to the user experience in which case we should not bother
Or it adds something, but then dropping to EDCA operation is unsatisfactory for users previously enjoying the fruits of 11aa’s labors
Raghuram Rangarajan, Cisco Systems

19. The simulations avoided the other hard problems related to OBSS
Mar 2009
The simulations avoided the other hard problems related to OBSS
Channel assignment is but one piece of the OBSS puzzle
Other known problems include
Communications overhead
The problems of sync (now “easy”)
Distributed scheduling (now “easier”)
Fairness (?)
Trust/security (tough across administrative domains)
non-AP STA relaying (PS?)
Raghuram Rangarajan, Cisco Systems

20. Mar 2009
The bottom line is that OBSS remains a hard problem without good answers
OBSS is hard and has been known to be so ever since an attempt was made to resolve it in TGe
Any complete solution is going to have to address more scenarios
And many of the other challenges that we are now ignoring
The TG should avoid jumping to a solution unless there is consensus that the solution is workable – this is a standardisation organisation after all and not a research organisation
Raghuram Rangarajan, Cisco Systems