fastand light bandwidth testing for internet users
TRANSCRIPT
![Page 1: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/1.jpg)
-1-
Fast and Light Bandwidth Testingfor Internet Users
Xinlei Yang , Xianlong Wang, Zhenhua Li, Yunhao Liu,Feng Qian, Liangyi Gong, Rui Miao, Tianyin Xu
![Page 2: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/2.jpg)
-2-
Outlines1. Background
2. Motivation
3. State-of-the-Art
4. Novel Design
5. Evaluation
6. System Demo
7. Conclusion
![Page 3: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/3.jpg)
-3-
pBandwidth testing services (BTSes) are widely used
1. Background
n Core component of many network applications
n Cited by government reports & trade press
n Handy measurement tools for Internet users
![Page 4: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/4.jpg)
-4-
1. BackgroundpBTSes are becoming increasingly important
n Virtual Network Operators (VNO) catching on
n Wireless access becoming ubiquitous
n Bandwidth-hungry apps (e.g., UHD videos, VR/AR) emerging
![Page 5: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/5.jpg)
-5-
2. Motivation
Example
BTSes Duration (s) Data Usage AccuracySpeedtest.att.com 19.1 1.37 GB 0.42Sourceforge.net 20.8 2.75 GB 0.81
Fast.com 13.5 1.20 GB 0.68SpeedTest.net 15.7 1.94 GB 0.87
mmWave 5G, 1.15-Gbps downlink bandwidth
pToday’s BTSes are not satisfactoryn Long test duration
n Excessive data usage
n Low accuracy for most BTSes
![Page 6: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/6.jpg)
-6-
2. Motivation
Example
BTSes Duration (s) Data Usage AccuracySpeedtest.att.com 19.1 1.37 GB 0.42Sourceforge.net 20.8 2.75 GB 0.81
Fast.com 13.5 1.20 GB 0.68SpeedTest.net 15.7 1.94 GB 0.87
mmWave 5G, 1.15-Gbps downlink bandwidth
pToday’s BTSes are not satisfactoryn Long test duration
n Excessive data usage
n Low accuracy for most BTSes
Can bandwidth testing be fast, light, and accurate simultaneously?
![Page 7: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/7.jpg)
-7-
Popular Bandwidth Testing Websites
Commercial Bandwidth Testing Apps
Important BandwidthTesting Interfaces
18 popular bandwidth testing websites+
WiFiMaster A popular Android/iOS appwith 800 million users
3. State-of-the-Art
Android 115G-oriented bandwidth testing Android SDK APIs
![Page 8: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/8.jpg)
-8-
3. State-of-the-ArtpResearch methodology
n Small-scale study
n Large-scale benchmarking
1. Network traffic tracing
2. System reverse engineering
![Page 9: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/9.jpg)
-9-
Setup Bandwidth Probing Bandwidth Estimation
3. State-of-the-ArtpSummarizing
TBB: thinkbroadband.com, SpeedOf: speedof.me, BWP: bandwidthplace.com, SFtest: sourceforge.net, ATTtest: Speedtest.att.com, Xfinity: speedtest.xfinity.com, FAST: fast.com, SpeedTest: speedtest.net, Android API-A: getLinkDownstreamBandwidthKbps()
![Page 10: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/10.jpg)
-10-
3. State-of-the-ArtpReflection of bandwidth testing
Time0
Thro
ughp
ut
Throughput samples
Time0
Thro
ughp
ut
Throughput samples
Ideal case Real case
Noises(congestion control,
link sharing, etc.)
User’s access link bandwidth
![Page 11: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/11.jpg)
-11-
3. State-of-the-ArtpCombating noises
Speedtest.net Fast.com
Using large-scale test server deployments (spatial redundancies) to ensure high-quality network connections, largely reducing noises.
Using long test duration (temporal redundancies) to wait for the coming of sufficient desired samples.
Test duration:often 20 - 30 s
Space Dimension Time Dimension
![Page 12: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/12.jpg)
-12-
Most of today’s BTSes use excessive temporal and spatial redundancies for combating noises
Large-scale network deployments,long test duration, and excessive data usage
Can we accommodate and exploit the noises rather than exhaustively suppress the impact of them?
4. Novel Design
![Page 13: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/13.jpg)
-13-
4. Novel Design
T(x)
Target Distribution
P(x)
Proposal Distribution
RejectedSamples
AcceptedSamples
Acceptance-RejectionFunction (ARF)
T(x)
pRe-consider BTS through rejection sampling
![Page 14: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/14.jpg)
-14-
Time0
Thro
ughp
ut
Throughput samples
4. Novel DesignpModeling the bandwidth testing process
ARFT(x)
Samples followingthe P(x) distribution
Bandwidth EstimationRejected samples
Accept
Reject
Accepted samples
Rejection Sampling
![Page 15: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/15.jpg)
-15-
Time0
Thro
ughp
ut
Throughput samples
4. Novel DesignpModeling the bandwidth testing process
ARFT(x)?
Samples followingthe P(x) distribution
Bandwidth EstimationRejected samples
Accept
Reject
Accepted samples
Fuzzy Rejection Sampling
![Page 16: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/16.jpg)
-16-
4. Crucial Interval Sampling (CIS)
Crucial Interval:(ARF)
Key Findings:Noise samples are scattered across a wide throughput interval
Desired samples tend to concentrate within a narrow interval
A throughput interval (1) whose density is as high as possible; and (2) which contains as many samples as possible
![Page 17: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/17.jpg)
-17-
Crucial intervals converge quickly
Accepted Samplesintensify the crucial interval by making it denser
Rejected Sampleshelp better “contrast” the crucial interval
Both accepted and rejected samples are exploited to make bandwidth tests fast and light.
4. Crucial Interval Sampling (CIS)
Video available at https://youtu.be/lgZOy59im7M
![Page 18: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/18.jpg)
-18-
O(N^2)
Convex hull acceleration O(NlogN)
4. Crucial Interval Sampling (CIS)pConvex hull acceleration
Brute-forcemechanism
Walking through all the throughput samples to find the crucial interval.
Dynamically maintaining a convex hull for quickly finding the crucial interval.
![Page 19: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/19.jpg)
-19-
4. Elastic Bandwidth Probing (EBP)Crucial interval not effective
n BBR: emerging congestion control mechanism with a built-in bandwidth probing scheme
n Leveraging and improving BBR to realize elastic bandwidth probing
n Making crucial interval always effective
![Page 20: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/20.jpg)
-20-
4. Novel DesignpArchitecture of FastBTS
Fuzzy Rejection Sampling
Crucial Interval Sampling
Elastic Bandwidth Probing
![Page 21: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/21.jpg)
-21-
4. Novel DesignpArchitecture of FastBTS
Fuzzy Rejection Sampling
Crucial Interval Sampling
Elastic Bandwidth Probing
Cannot saturate user’s access link bandwidth
0 0.2 0.4 0.6 0.8 1
Accuracy0
0.2
0.4
0.6
0.8
1
CD
F
SequentialProgressive ParallelFixed-number ParallelFixed-number Multi-sourceProgressive Multi-source
Data-driven Server Selection
Adaptive Multi-homing
![Page 22: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/22.jpg)
-22-
4. Data-driven Sever Selection (DSS)
Ping-basedserver selection
Historical performance-based sever selection
Low latency ≠ high throughput
Select candidateservers
Crucial latency interval
HistoricalData max
avg Sorting servers
Select servers with highestbandwidth estimations
Ping test
![Page 23: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/23.jpg)
-23-
4. Adaptive Multi-Homing (AMH)
Adding concurrency level with fixed threshold
Adaptive Multi-Homing
Under-estimating user’s bandwidth (e.g., 5G)
When shall we stopadding extra test servers?
Adding concurrency level
Aggregatedthroughputincreasing?
Generatingresult
N
Y
![Page 24: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/24.jpg)
-24-
5. EvaluationpTestbed networks
Residential broadband Datacenter networkLAN
LTE network mmWave & Sub-6Ghz 5G network HSR cellular network
![Page 25: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/25.jpg)
-25-
5. Evaluation
n FastBTS vs. others on testbed networks: 5%–72% higher average accuracy,
2.3–8.5× shorter test duration, 3.7–14.2× less data usage.
n FastBTS vs. SpeedTest.net in real world: FastBTS (with only 30 servers)
achieves comparable accuracy compared with the production system of
SpeedTest.net with ~12,000 test servers, incurring 5.6× shorter test
duration and 10.7× less data usage on average.
pMajor results
![Page 26: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/26.jpg)
-26-
6. System DemopCase 1 : PC + Wi-Fi (~100 Mbps)
SpeedTest.net FastBTS.thucloud.com
Duration: 15.0 secondsResult: 95.18 MbpsData usage: 176 MB
Duration: 3.1 secondsResult: 99.25 MbpsData usage: 37 MB
Videos available at: https://youtu.be/QbHO27RvzbU
![Page 27: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/27.jpg)
-27-
6. System DemopCase 2 : smartphone + Sub-6Ghz 5G (~500 Mbps)
SpeedTest.net FastBTS.thucloud.com
Duration: 15.0 secondsResult: 484 MbpsData usage: 936 MB
Duration: 4.1 secondsResult: 543.07 MbpsData usage: 168 MB
Videos available at: https://youtu.be/VGN32d3dIAU
![Page 28: Fastand Light Bandwidth Testing for Internet Users](https://reader030.vdocuments.us/reader030/viewer/2022012609/619c44e21bd33e410c453811/html5/thumbnails/28.jpg)
-28-
n We reveal how today’s commercial bandwidth testing services actually work as well as
their pros and cons based on in-depth investigations and large-scale benchmarking tests.
n We present FastBTS, a novel bandwidth testing solution that accommodates and
exploits network noises to make bandwidth tests fast and light. With only 30 test servers,
FastBTS achieves comparable accuracy compared with SpeedTest.net with ~12,000
servers, while incurring 5.6× shorter test duration and 10.7× less data usage on average.
n We have released all the source code at https://FastBTS.github.io and an online demo
system at http://FastBTS.thucloud.com.
7. Conclusion
Thanks!Q & A