large-scale measurements of wireless network...
TRANSCRIPT
Large-scale Measurements of Wireless Network BehaviorSanjit Biswas ([email protected]), John Bicket ([email protected]), Edmund L. Wong ([email protected]), Raluca Musaloiu-E ([email protected]), Apurv Bhartia ([email protected]), Dan Aguayo ([email protected])
Over 30 years of unlicensed devices
2
• Over a billion new devices added each year – phones, laptops, security cameras, headsets, fitness trackers
Source: http://www.ce.org/CorporateSite/media/gla/CEAUnlicensedSpectrum WhitePaper-FINAL-052814.pdf
1985: ISM band
created
1989: New power
limits
1999: 802.11a
1998: 802.11b
1999: Bluetooth
1999: U-NII
2002: 802.11g
2009: 802.11n
How does unlicensed spectrum perform in the real world?
Billions of cellular devicesControlled channel access,
High power limit (3000W/MHz),
Billions of unlicensed devices
Open channel access,Low power limit (1W total),
802.11 works well in the lab – what about the “real world”?
3
Outline• How our measurements were collected
• What does the application and device workload look like?
• How prevalent is interference?
• Lessons and conclusions
4
Meraki’s unique perspective
• Centralized management platform for thousands of networks worldwide• Time-series data of application, client and device statistics
5
This study: data from thousands of networks
• Collected data over a one week period from 20,667 networks, in Jan. 2014 and Jan. 2015
• 5.58M clients across wide variety of deployment types
6
Industry # networks Architecture/Engineering 127 Construction 333 Consulting 365 Education 4,075 Finance/Insurance 737 Government/Public Sector 1,112 Healthcare 1,382 Hospitality 493 Industrial/Manufacturing 1,220 Legal 264 Media/Advertising 427 Non-Profit 640 Real Estate 386 Restaurants 296 Retail 2,355 Tech 983 Telecom 442 VAR/System Integrator 2,876 Other 2,154 Total 20,667
7
How are people using WiFi in 2015?
Usage by application
• Media accounts for large fraction of traffic (YouTube, Netflix, iTunes)• Heavy networks may traffic shape – bitrate adaptive, latency insensitive traffic
8
Application Category TB (% total) % increase # clients MB /
client Miscellaneous web Other 239 (13%) 67% 4,623,630 54 YouTube Video & music 202 (11%) 93% 1,934,371 110 Netflix Video & music 188 (9.8%) 76% 161,014 1,224 Non-web TCP Other 156 (8.2%) 51% 3,656,494 45
Miscellaneous secure web Other 147 (7.7%) 94% 5,115,023 30
iTunes Video & music 102 (5.4%) 66% 2,230,787 48 Miscellaneous video Video & music 98 (5.1%) 61% 1,383,386 74 Windows file sharing File sharing 87 (4.5%) 48% 740,591 123 CDNs Other 75 (3.9%) 81% 3,157,028 25 UDP Other 61 (3.2%) 60% 3,705,171 17
Facebook Social 53 (2.8%) 127% 3,579,926 16
Usage by device type
• Smartphones outnumber laptops 4:1, but usage on laptops is much higher• Average usage per device growing faster on smartphones
9
OS TB (% total) % increase # clients % increase MB / client % increase
Windows 589 (30%) 43% 822,761 28% 751 12% Apple iOS 545 (28%) 92% 2,550,379 34% 224 44% Mac OS X 445 (23%) 44% 313,976 24% 1,487 17%
Android 177 (9.1%) 172% 1,535,859 61% 121 69%
Unknown 78 (4.0%) -9.2% 228,182 -8.9% 357 -0.36% Chrome OS 62 (3.2%) 275% 178,095 222% 366 16%
All 1,950 (100%) 62% 5,578,126 37% 367 18%
What types of radios are out there?• Moore’s law advances in signal processing
have led to more streams, higher bandwidths
• However, 802.11n does not imply MIMO– iPhone 5, iPhone 6 use single stream 11n/ac– Due to antenna space, power, “fast enough”
• Implication for protocol design: – New technologies take time to have an impact– Co-existence is common, even with 5+ year
old standards
10
Jan. 2014 Jan. 2015
802.11g 99.9% 99.9% 802.11n 95.7% 97.7% 5 GHz 48.9% 64.9% 40 MHz channels 23.4% 63.8% 802.11ac 2.5% 18.0% Two streams 7.7% 19.3% Three streams 2.4% 3.8% Four streams 0.7% 1.8%
11
Interference
Hardware platform• Full control of hardware and firmware,
along with identical hardware makes it easy to look at data
• Meraki MR16: older design with dual concurrent radios
• Meraki MR18: newer design with scanning radio for spectrum analysis– 2x2 2.4GHz 802.11n – 2x2 5GHz 802.11n– 1x1 2.4/5GHz scanning radio
12
Nearby networks
• Sudden increase in 2014: personal hotspots, guest access SSIDs• Difficult to measure changes in client throughput, but inter-AP link delivery rates in
the 2.4GHz band are 10% lower 13
Networks
Networks per AP
2.4 GHz (now) 527,087 55.472.4 GHz (six months ago) 230,628 28.60 5 GHz (now) 35,010 3.68
5 GHz (six months ago) 19,921 2.47
0 10 20 30 40 50 60 70 80 90 100
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
2.4 GHz − 6 months ago2.4 GHz − now5 GHz − 6 months ago5 GHz − now
Delivery ratio
Cum
ulat
ive fr
actio
n of
link
s
Channel utilization
• Energy detect is triggered 20% of the time for the median 2.4GHz radio, 50% for top 10th percentile
• Nearby network count doesn’t predict utilization – better to measure directly14
0
0.2
0.4
0.6
0.8
1
0 20 40 60 80 100
Cum
ulat
ive
Frac
tion
ofC
hann
el U
tiliz
atio
n M
easu
rem
ents
Channel Utilization
5GHz (PM)
5GHz (AM)
2.4GHz (PM)
2.4GHz (AM) 0
5
10
15
20
25
30
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Num
ber o
f Int
erfe
ring
APs
Band Utilization (%)
What is causing the interference?
• Much of the channel utilization caused by other 802.11 traffic• Helpful for MAC protocol design to be able to decode headers
15
0
0.2
0.4
0.6
0.8
1
0 20 40 60 80 100
Cum
ulat
ive
Frac
tion
of T
otal
Util
izat
ion
Mea
sure
men
ts
Decodable Traffic (%)
2.4GHz
5GHz
Related work• Link measurements
– [Aguayo04]: data from 38-node outdoor 802.11b network with intermediate links– [Reis06], [Halperin11]: data from 15 node indoor network showing selective fading– [Gollakota08]: effects of interference and cancellation methods
• Network studies– [Ghosh11] study of AT&T’s hotspot network with 240k client devices– [Afanasyev10] Google WiFi’s 500 node outdoor network with 30k client devices– [Gember11] study of UWisc network with 32k devices
• This paper’s primary contributions are measurements across many networks and a look at real-world interference
16
Lessons and conclusions• At large scale, you see extremes of distributions
– Access points with 10,000 nearby networks (passenger bus in Manhattan)– Cat 5/6 cable problems in large networks– Overloaded RADIUS servers causing client device auth problems
• Network design needs to adapt over time– In 2006, very few smartphones in use – now they are majority of devices– Rethink assumptions around device roaming, addressing and hardware– Application workloads shifted from web to video – traffic shaping more useful
• Network operators and protocol designers should assume significant interference from legacy devices – too many out there to ignore
17
Thanks!
18
Anonymized dataset available at:http://dl.meraki.net/sigcomm-2015