Self ManagementSelf Managementin Chaotic Wireless in Chaotic Wireless

DeploymentsDeployments

2

This paper …This paper …• Was supported by the army research Was supported by the army research

office, NSF, Intel and IBM.office, NSF, Intel and IBM.

• Characterizes the density and usage of Characterizes the density and usage of 802.11 hardware across major US cities.802.11 hardware across major US cities.

• Presents a simulation study of the effect Presents a simulation study of the effect of dense unmanaged wireless of dense unmanaged wireless deployments on end-user performancedeployments on end-user performance

3

This paper…This paper…

• Outlines the challenges to make Outlines the challenges to make chaotic environments self-managingchaotic environments self-managing

• Describes algorithms to increase the Describes algorithms to increase the quality of performance.quality of performance.

• Examines these algorithms.Examines these algorithms.

4

Problems with WiFi Problems with WiFi networksnetworks

• Wireless links are susceptible to Wireless links are susceptible to degredation (fading)degredation (fading)

• Sharing scarce spectrum by wireless Sharing scarce spectrum by wireless deployments causes interference. deployments causes interference.

• Problems due to wide usage of hardware Problems due to wide usage of hardware in an unmanaged and unplanned manner.in an unmanaged and unplanned manner.

5

Chaotic deploymentChaotic deployment

• To give you an idea, 4.5 million WiFi To give you an idea, 4.5 million WiFi APs were sold in 3APs were sold in 3rdrd quarter of 2004 quarter of 2004 and will triple by 2009. and will triple by 2009.

• Unplanned (Not planned to optimize Unplanned (Not planned to optimize the coverage, spontaneous deployment)the coverage, spontaneous deployment)

• Unmanaged ( Not configured to have Unmanaged ( Not configured to have the right parameters)the right parameters)

6

Chaotic deployments Chaotic deployments bring…bring…

• Serious contentionSerious contention

• Poor performancePoor performance

• Security risksSecurity risks

• The main goal of the paper is to show the The main goal of the paper is to show the effects of interference on performance.effects of interference on performance.

7

Related WorkRelated Work

• Evaluation based on current efforts Evaluation based on current efforts to map 802.11 deploymentsto map 802.11 deployments

• Overview of commercial products for Overview of commercial products for managing wireless networksmanaging wireless networks

• Proposal for wireless self Proposal for wireless self management and examining those.management and examining those.

8

• Several internet websites map WiFi hot-Several internet websites map WiFi hot-spots in different US cities.spots in different US cities.

• WiFimaps.com, Wi-Fi-Zones.com, WiFimaps.com, Wi-Fi-Zones.com, JIWire.comJIWire.com

• Data from wifimaps.com and intel place Data from wifimaps.com and intel place lab database is used to infer usage lab database is used to infer usage characteristics.characteristics.

9

Data SetsData Sets

10

Measurement Measurement ObservationsObservations

• Focus on using the wireless spectrum Focus on using the wireless spectrum efficiently by developing algorithms in efficiently by developing algorithms in dense wireless networks not saving energydense wireless networks not saving energy

• Not complete data sets due to increasing Not complete data sets due to increasing rate of wireless networks and density.rate of wireless networks and density.

• Information about other devices using the Information about other devices using the same spectrum is not gathered and shown same spectrum is not gathered and shown

11

Deployment DensityDeployment Densityplace lab data setplace lab data set

• Interference range assumed 50mInterference range assumed 50m

• Two nodes are neighbors if in each other’s interference rangeTwo nodes are neighbors if in each other’s interference range

• In most cities, the degree of APs is 3.In most cities, the degree of APs is 3.

• Table shows the close proximity and density of wireless networks.Table shows the close proximity and density of wireless networks.

12

ChannelsChannels

• Information here Information here suggests that most suggests that most APs that overlap in APs that overlap in coverage are not coverage are not configured to configured to optimize optimize performance by performance by minimizing minimizing interference.interference.

13

Vendors and AP Vendors and AP Management SupportManagement Support

• If Linksys and If Linksys and Aironet incorporate Aironet incorporate built-in self built-in self management management firmware, we will firmware, we will see a sharp see a sharp decrease in negative decrease in negative impacts of impacts of interference in interference in chaotic chaotic deployments.deployments.

14

Impact on End-User Impact on End-User performanceperformance

• Assumptions made as following:Assumptions made as following:

Each node is an APEach node is an AP Each node has D clients (0 < D < 4)Each node has D clients (0 < D < 4) Clients are located less than 1m from APClients are located less than 1m from AP Transmission done on channel 6Transmission done on channel 6 Fixed transmit power level of 15dBmFixed transmit power level of 15dBm Transmit rate is the same for all at 2MbpsTransmit rate is the same for all at 2Mbps RTS/CTS is turned off.RTS/CTS is turned off. Stretch: the higher it is, the lower the impact of Stretch: the higher it is, the lower the impact of

interference.interference.

15

Cont…Cont…

16

Cont…Cont…

• The impact of interference in chaotic The impact of interference in chaotic deployments depends largely on users deployments depends largely on users workloads.workloads.

• Very likely to not experience any Very likely to not experience any degradation when transmitting data degradation when transmitting data occasionally occasionally

• The goal of this evaluation is to The goal of this evaluation is to quantify the exact impact of user quantify the exact impact of user workloadworkload

17

Set of WorkloadsSet of Workloads• The first set if HTTP.The first set if HTTP.• There is a think time on client side There is a think time on client side

between each HTTP transfer (s seconds).between each HTTP transfer (s seconds).• Authors vary s between values 5 and 20s.Authors vary s between values 5 and 20s.• Average loads:Average loads:

– 83.3Kbps for 5s sleep time83.3Kbps for 5s sleep time– 24.5Kbps for 20s sleep time24.5Kbps for 20s sleep time

• No other interfering traffic than HTTPNo other interfering traffic than HTTP

18

Cont…Cont…

• The second set is FTP called comb-ftpThe second set is FTP called comb-ftpii..• i clients running long-lived FTP traffic.i clients running long-lived FTP traffic.• 0 < i < 40 < i < 4• Average load is 0.89Mbps.Average load is 0.89Mbps.

Each set of workloads run for about Each set of workloads run for about 300s300s

19

Interference at Low Client Interference at Low Client Densities and Traffic Densities and Traffic

VolumesVolumes• Impact of interference Impact of interference

under light-weight user under light-weight user traffic on each AP and traffic on each AP and low client density (D = low client density (D = 1).1).

• Normalized Normalized performance is the ratio performance is the ratio of average throughput of average throughput flow to the throughput flow to the throughput when operating in when operating in isolation.isolation.

20

ResultsResults• Performance of HTTP improves until Performance of HTTP improves until

stretch 10.stretch 10.• After stretch 10 the behavior stays the After stretch 10 the behavior stays the

same.same.• When HTTP component is aggressive When HTTP component is aggressive

(s= 5s), FTP suffers by 17%.(s= 5s), FTP suffers by 17%.• When HTTP not aggressive, the When HTTP not aggressive, the

impact on FTP is minimal. impact on FTP is minimal.

21

Cont…Cont…

• Impact of Impact of interference with interference with light-weight traffic light-weight traffic but high user but high user density (D = 3).density (D = 3).

22

ResultsResults

• Performance of both HTTP and FTP Performance of both HTTP and FTP suffers significantly under high client suffers significantly under high client density.density.

• In figure 4a, HTTP and FTP performance In figure 4a, HTTP and FTP performance decrease about 65% with s = 5s.decrease about 65% with s = 5s.

• When s = 20 s, HTTP suffers by 20% and When s = 20 s, HTTP suffers by 20% and FTP performance is lowered by 36%.FTP performance is lowered by 36%.

23

Cont…Cont…

• Impact of higher Impact of higher traffic loads and traffic loads and client density, client density, Comb-ftp{2,3}, Comb-ftp{2,3}, D = 3D = 3

• The impact on The impact on performance is performance is sensiblesensible

24

Limiting the Impact of Limiting the Impact of Interference Interference

• Two goalsTwo goals

If Optimal static non-overlapping If Optimal static non-overlapping channel allocation eliminates channel allocation eliminates interference altogether?interference altogether?

Preliminary investigation of the effect of Preliminary investigation of the effect of reducing transmit power levels at reducing transmit power levels at access points on interference.access points on interference.

25

Optimal Static Channel Optimal Static Channel AllocationAllocation

• Impact of static Impact of static channel allocation, channel allocation, set to the three non-set to the three non-overlapping channelsoverlapping channels

• Transmit power level Transmit power level set 15dBm set 15dBm corresponding to corresponding to 31m31m

26

ResultsResults

• Curves flatten out because of optimal static channel Curves flatten out because of optimal static channel allocation.allocation.

• There is still poor performance.There is still poor performance.

• HTTP is performing 25% lower at stretch 1 HTTP is performing 25% lower at stretch 1 comparing to stretch 10.comparing to stretch 10.

• FTP performance is still suffering.FTP performance is still suffering.

• Optimal channel allocation can not eliminate the Optimal channel allocation can not eliminate the interference entirely.interference entirely.

27

Transmit Power ControlTransmit Power Control

• Optimal static Optimal static channel along with channel along with setting transmit setting transmit power level at power level at 3dBm 3dBm corresponding to corresponding to 15m.15m.

28

Results Results

• The overall performance improves The overall performance improves significantly.significantly.

• At stretch 1, all HTTP and comb-ftp traffic is At stretch 1, all HTTP and comb-ftp traffic is doing much better, about 20% more.doing much better, about 20% more.

• At stretch 2 the curve flattens out. At stretch 2 the curve flattens out.

• This experiment shows that transmit power This experiment shows that transmit power control along with optimal channel allocation control along with optimal channel allocation could reduce the impact in chaotic networks.could reduce the impact in chaotic networks.

29

Improvement in Network Improvement in Network CapacityCapacity

• D = 1 D = 1

30

ResultsResults

• The capacity of a densely packed network of APs is The capacity of a densely packed network of APs is 15% of the maximum capacity.15% of the maximum capacity.

• Static channel allocation helps the capacity two-Static channel allocation helps the capacity two-fold.fold.

• Lower transmit power on APs improves capacity Lower transmit power on APs improves capacity by nearly a factor of 2.by nearly a factor of 2.

• Transmit power control along with optimal Transmit power control along with optimal channel allocation ensures a much better channel allocation ensures a much better performance.performance.

31

Benefits of Transmit Power Benefits of Transmit Power ReductionReduction

• Assumptions made:Assumptions made:

Consider only Consider only downlink traffic as downlink traffic as uplink traffic is small.uplink traffic is small.

The algorithm works The algorithm works for both.for both.

Each AP has a single Each AP has a single client at a fixed client at a fixed distance.distance.

32

• The minimum physical spacing The minimum physical spacing between APs is important.between APs is important.

• First the medium utilization is First the medium utilization is computed:computed:

Util Util ap ap = load / throughput = load / throughput maxmax

Pathloss = 40+3.5*10* log( dPathloss = 40+3.5*10* log( dclientclient))RSS = txPOWER – pathlossRSS = txPOWER – pathlossSNR = -100SNR = -100

33

ResultsResults

34

• After computing utilization for a single link by the After computing utilization for a single link by the formula above, the minimum utilization is formula above, the minimum utilization is computed by summing utilization of all in range computed by summing utilization of all in range APs.APs.

• Two APs are in range if Two APs are in range if RSS > interference threshold RSS > interference threshold

• For this paper : interference threshold = -100For this paper : interference threshold = -100

• Next graph shows the results for a client distance Next graph shows the results for a client distance of 10m and loads ranging from 0.1 Mbps to 1.1 of 10m and loads ranging from 0.1 Mbps to 1.1 Mbps.Mbps.

35

Resulting GraphResulting Graph

36

Conclusions from GraphConclusions from Graph

• The minimum distance between APs The minimum distance between APs decreases( higher density)decreases( higher density)

• By lowering the transmit power rate, denser By lowering the transmit power rate, denser networks can be deployed.networks can be deployed.

• the upper bound of power is in hand (x-axis).the upper bound of power is in hand (x-axis).

• When the load is not high, the node can When the load is not high, the node can reduce both transmit rate and power in order reduce both transmit rate and power in order to increase network capacity. to increase network capacity.

37

Deployment ChallengesDeployment Challenges• There is a trade-off when using these techniques and There is a trade-off when using these techniques and

that is a reduction in throughput of the channel by that is a reduction in throughput of the channel by forcing the transmitter to use a lower rate to deal forcing the transmitter to use a lower rate to deal with the reduced signal to ratio.with the reduced signal to ratio.

• Determining the right moment and environment to Determining the right moment and environment to use them can greatly affect the network.use them can greatly affect the network.

• There is a trade-off between selfish and social There is a trade-off between selfish and social congestion control in the internet.congestion control in the internet.

• One advantage is that lower transmission rate limits One advantage is that lower transmission rate limits the chances of eavesdropping for malicious the chances of eavesdropping for malicious attackers. attackers.

38

Transmission Power and Transmission Power and Rate Selection Rate Selection

• A prism chipset 2.5 NIC card is used.A prism chipset 2.5 NIC card is used.

• Driver is a modified version of hostAP prism driver in Linux.Driver is a modified version of hostAP prism driver in Linux.

• The driver achieves per packet control over transmission The driver achieves per packet control over transmission rate by tagging each packet with the rate at which it should rate by tagging each packet with the rate at which it should be sent ( retransmission is set to 2)be sent ( retransmission is set to 2)

• Prism based cards do not support per packet transmission Prism based cards do not support per packet transmission power control ( prism 2.5 firmware)power control ( prism 2.5 firmware)

• Overcome to this limitation is to wait for the NIC buffers to Overcome to this limitation is to wait for the NIC buffers to empty and then queue packets at a new rate.empty and then queue packets at a new rate.

39

Fixed Power Rate Selection Fixed Power Rate Selection AlgorithmsAlgorithms

• ARFARF : Auto Rate Fallback ( mostly : Auto Rate Fallback ( mostly used)used)

• ERFERF: Estimated Rate Fallback: Estimated Rate Fallback

40

Auto Rate FallbackAuto Rate Fallback

• ARF attempts to choose the best transmission rate ARF attempts to choose the best transmission rate via in-band probing using 802.11’s ACK mechanism.via in-band probing using 802.11’s ACK mechanism.

• There are variations of ARF. The one used in 802.11b There are variations of ARF. The one used in 802.11b assumes the following:assumes the following:

– Failed transmission indicates a very high transmission rate.Failed transmission indicates a very high transmission rate.

– Successful transmission indicates a good transmission rate Successful transmission indicates a good transmission rate and that a higher one is possible.and that a higher one is possible.

– An increment threshold of 6, decrement threshold of 3 and An increment threshold of 6, decrement threshold of 3 and 10s idle timout.10s idle timout.

41

Proposed ARF AlgorithmProposed ARF Algorithm

• The authors make modifications to ARF as The authors make modifications to ARF as following if a threshold number of consecutive following if a threshold number of consecutive packets are:packets are:

– sent successfully, the node chooses the next rate.sent successfully, the node chooses the next rate.

– Not sent successfully, the node decrements the rate.Not sent successfully, the node decrements the rate.

– Dropped entirely, the highest rate is chosen.Dropped entirely, the highest rate is chosen.

– Thresholds are set to 6 successful, 4 failed and 10s Thresholds are set to 6 successful, 4 failed and 10s for idle timeout. for idle timeout.

42

SNR, Alternative for ARF SNR, Alternative for ARF and Challengesand Challenges

• Advantage:Advantage:

– Using channel’s SNRto select the optimal rate for a Using channel’s SNRto select the optimal rate for a given SNR instead of probing channels for best rate.given SNR instead of probing channels for best rate.

• Disadvantage:Disadvantage:

– Card measurements of SNR can be inaccurate and Card measurements of SNR can be inaccurate and may vary between different cards.may vary between different cards.

– SNR measurement can completely identify channel SNR measurement can completely identify channel degradation due to multi-path interference.degradation due to multi-path interference.

43

Proposed ERF AlgorithmProposed ERF Algorithm• SNR based algorithm.SNR based algorithm.

• A hybrid between SNR and ARF.A hybrid between SNR and ARF.

• Uses path-loss information to estimate the SNR with which Uses path-loss information to estimate the SNR with which transmission is received.transmission is received.

• ERF then determines the highest rate supportable by this ERF then determines the highest rate supportable by this SNR.SNR.

• In case of a successful or unsuccessful attempt, it increments In case of a successful or unsuccessful attempt, it increments or decrements the rate.or decrements the rate.

• If all packets are dropped, ERF will begin to fall back If all packets are dropped, ERF will begin to fall back towards the lowest rate until new channel info is received.towards the lowest rate until new channel info is received.

44

PARF and PERFPARF and PERF• Both algorithms try to reduce the transmission Both algorithms try to reduce the transmission

power. (social reduction interference)power. (social reduction interference)

• At highest rate, PARF reduces the power after At highest rate, PARF reduces the power after successful operation.successful operation.

• Repeats the process until the lowest power is Repeats the process until the lowest power is reached or fails.reached or fails.

• Then the power is raised until no fails occur.Then the power is raised until no fails occur.

45

Cont…Cont…• For PERF, an estimated SNR is computed at For PERF, an estimated SNR is computed at

the receiver.the receiver.

• If ESNR is higher than decision threshold for If ESNR is higher than decision threshold for the highest rate then transmit is reduced until the highest rate then transmit is reduced until

ESNR = decisionThreshold + ESNR = decisionThreshold + powerMarginpowerMargin

• powerMargin allows Aggressiveness of power powerMargin allows Aggressiveness of power control algorithm to be tuned.control algorithm to be tuned.

46

Performance Evaluation Performance Evaluation and Conclusion and Conclusion

47

Cont…Cont…

• PARF shows unstable in initial experiment.PARF shows unstable in initial experiment.

• PERF reacts more slowly to transmission failure. PERF reacts more slowly to transmission failure.

• Poor performance of ERF and ARF because of asymmetric Poor performance of ERF and ARF because of asymmetric carrier sense. carrier sense.

• The authors also introduce one strategy to improve PERF The authors also introduce one strategy to improve PERF called LPERF (load-sensitive PERF) in which transmitters called LPERF (load-sensitive PERF) in which transmitters reduce their power even if it reduces their transmission reduce their power even if it reduces their transmission rate.rate.

• They claim that LPERF like algorithms are a promising They claim that LPERF like algorithms are a promising direction.direction.