Factors Affecting Performance of Web Flows in Cellular Networks - - PowerPoint PPT Presentation

factors affecting performance of web flows in cellular
SMART_READER_LITE
LIVE PREVIEW

Factors Affecting Performance of Web Flows in Cellular Networks - - PowerPoint PPT Presentation

Factors Affecting Performance of Web Flows in Cellular Networks Ermias A. Walelgne , Kim Set al a, Vaibhav Bajpai, Stefan Neumeier, Jukka Manner, J org Ott May 15, 2018 - IFIP Networking, Zurich Introduction Introduction


slide-1
SLIDE 1

Factors Affecting Performance of Web Flows in Cellular Networks

Ermias A. Walelgne , Kim Set¨ al¨ a, Vaibhav Bajpai, Stefan Neumeier, Jukka Manner, J¨

  • rg Ott

May 15, 2018 - IFIP Networking, Zurich

slide-2
SLIDE 2

Introduction

slide-3
SLIDE 3

Introduction

Introduction — Motivation

  • ∼ 99% of the Internet traffic flows are short (<100 KB).

[ Brownlee and claffy SIGMETRICS’02 , Ramachandran Google’10]

  • > 95% traffic generated by smartphones consists short-lived TCP flows.

[Huang et al. SIGCOMM’13]

Performance of short web flows driven by latency than network through- put: DNS lookup time TCP connect time

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 1 / 21

slide-4
SLIDE 4

Introduction

Introduction — Research Question

Few studies that quantify the factors that are responsible for DNS lookup & TCP connect times in cellular network. [Xu et al. SIGMETRICS’11, Rula and Bustamante, SIGCOMM’14] We want to know: What are factors affecting DNS lookup and TCP connect time? How much DNS cached entries and TCP proxies improve latency? Distribution of packet loss and DNS look up failure.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 2 / 21

slide-5
SLIDE 5

Introduction

Introduction — Contribution

1

DNS lookup failure & packet loss

∼ 2% DNS lookup test experience failures ∼ 14.98% of have lost at least one packet

2

Radio technology & device model:

TCP connect times to popular websites are reduced by ∼80% on LTE compared to legacy networks. Device model has an impact on DNS lookup time.

3

ISP caches & DNS server’s proximity:

ISP caches improve TCP connect times towards some websites. DNS server’s proximity to the subscriber has an impact on DNS lookup time.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 3 / 21

slide-6
SLIDE 6

Methodology

slide-7
SLIDE 7

Methodology

Measured Websites

DNS lookup and TCP Connect time towards 4 websites: www.google.fi www.youtube.com www.facebook.com www.elisa.net Ping Test towards: www.google.fi

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 4 / 21

slide-8
SLIDE 8

Methodology

Measurement —DNS Lookup Time — TCP Connect Time — Ping Test

DNS Lookup Time:

DNS lookup time (in milliseconds) IPv4 address of DNS server Radio technology, device model Response error code

TCP Connect Time

Starting time of the test FQDN of the destination host Radio technology, device model

Ping Test:

ICMP echo request towards www.google.fi RTT and packet loss five to nine ICMP Echo requests payload size of request is 16 bytes

DNS lookup test

Measures the time it takes to look up a FQDN from a DNS server

TCP connect time

Measures the time to connect to a target website ( IPv4 ,80 ) from the client

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 5 / 21

slide-9
SLIDE 9

Methodology

Measurement Setup

Measurements tests are executed inside sessions. A session starts when a network interface becomes available or best interface changes. It is not periodic, but they are repeated when network conditions changes

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 6 / 21

slide-10
SLIDE 10

Methodology

Data Set and Measurement Trials The geographical distribution of ∼25K subscribers in Finland. Website DNS (#) TCP (#) ping (#) www.facebook.com 3.4M 4.6M

  • www.google.fi

6.9M 4.9M 2.1M www.youtube.com 1.6M 4.1M

  • www.elisa.net

1.8M 5.3M

  • DNS, TCP and ping measurements by

website. A month-long dataset collected through a mobile operator in Finland (Elisa)

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 7 / 21

slide-11
SLIDE 11

Data Analysis & Results

slide-12
SLIDE 12

Data Analysis & Results

DNS Lookup Failures

FacebookGoogle Youtube Elisa 1 2 3 2.16 0.96 2.99 2.74 % Failures per website (LTE ) ∼ 2% of the total DNS lookup show failure ∼ 86% of the DNS failures indicating that a responder does not implement the version level of the request LTE (1.9%) , UMTS (3.4%) , HSPA (3.9%) and HSPA+ ( 2.7%)

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 8 / 21

slide-13
SLIDE 13

Data Analysis & Results

Latency — Using Ping Test

Min, Avg & Max RTT values split by radio technology for a ping towards www.google.fi

∼ 90% of the average ping test towards www.google.fi using LTE have a RTT < 100 ms. Legacy 3G technologies are quite slow with more than 200 ms RTT.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 9 / 21

slide-14
SLIDE 14

Data Analysis & Results

Ping Test — packet loss by radio technology

Distribution of packet loss as the fraction total ping by radio technology type.

Of all ping tests over LTE , 2.4% of them lost at least a single packet. ping test over UMTS network experience highest packet loss (∼ 65%).

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 10 / 21

slide-15
SLIDE 15

Data Analysis & Results

Ping Test — packet loss by # packets sent at every ping test instance

Percentage of packets loss across the number of packets sent.

∼14.98% of tests in ping measurement have at least one packet loss. Packet loss happens, if the number of packets sent at every ping test instance > 5 Echo Requests.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 11 / 21

slide-16
SLIDE 16

Data Analysis & Results

DNS lookup time — by radio technology

LTE exhibits significantly lower latency. 75% www.youtube.com < 200ms [LTE] 25% www.youtube.com < 200ms [3G]

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 12 / 21

slide-17
SLIDE 17

Data Analysis & Results

TCP connect time — by radio technology

TCP connect time towards www.youtube.com (L) & www.google.fi (R)

.

TCP Connect time towards www.youtube.com 92% of TCP test using LTE finished < 100ms

  • nly 28% of 3G based TCP test finished < 100ms

The distribution exhibits similar pattern for www.elisa.net & www.facebook.com.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 13 / 21

slide-18
SLIDE 18

Data Analysis & Results

DNS lookup time — Device models

DNS response time of www.google.fi (L) and www.facebook.com (R) across device models as measured over LTE − order by device models’ release year.

No clear pattern between DNS lookup time & device models year of release Variation in DNS resolution time among device models is very high Google has faster resolution time in most devices than Facebook (median case)

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 14 / 21

slide-19
SLIDE 19

Data Analysis & Results

TCP connect time — Device models

.-/ eo

,T 103

a-

E

+-( +-(

a- C C 0 102 a`

I. 101

I

  • I

1 I 2

I

  • I
  • 3

I

  • I
  • I
  • I

4

  • :

I

. I I

I

I I

  • :

I

  • 5

6

I

  • 7
  • I

I

  • I
  • 8

9

I I i : I I

Devjce rfleased year :

I . . I I

  • 10 11

: ! I

12

I I

  • 13 14 15

I

  • 2012
I
  • I
  • 16 17

I I I I ! :

I :

  • 2013. 2014. 2015

.$ 2?16;

  • 18 19

2 2 1 22

  • 23
  • I
  • 24

I

I

I

25 Device Model Identifier

I

  • i

2 6

  • I
  • I

27

  • I
  • 2

8

  • I

I

  • I
  • 29

I

I

I I

I I I

30

  • I
  • I
  • 104
  • I
  • I
  • I
  • I
  • I
  • I

I I

  • I
  • .-/
  • eo
  • E

I

  • 103

I

a-

  • E

I

+-( +-(

a- C C 0 102 a` I.

  • I

101

  • 1

2 3 4 5 6 7 8 9

I I

  • I
  • Devjce rfleased year :

I

i

I

  • I •
  • I
  • I
  • I

I

I

  • I
  • I
  • I
  • I
  • I
  • I

I

  • I •
  • 10 11

12 13 14 15

  • 2012
  • I
  • I

I I

  • I
  • I
  • 16 17

I I I I •

I

2013-2014-2015 $ 2016 •

  • I

I I I

  • I

I

  • 18 19

2 2 1 22 23 24 25 Device Model Identifier

  • I •
  • I
  • I
  • I
  • I

I

  • I

I 2 6 27 2 8 29 30

TCP connect time for www.google.fi (L) and www.facebook.com (R) across device models as measured over LTE − order by device models’ release year.

Device type has smaller impact to TCP connect time Both Google and Facebook have similar TCP Connect time for most of device models

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 15 / 21

slide-20
SLIDE 20

Data Analysis & Results

DNS lookup time — Websites

DNS response time towards websites using LTE − towards different DNS resolvers.

DNS server’s proximity to the subscriber has an impact on DNS lookup time. www.youtube.com and www.facebook.com are slower than www.google.fi (likely cached by DNS resolvers) & www.elisa.net (ISP’s website).

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 16 / 21

slide-21
SLIDE 21

Data Analysis & Results

TCP connect time — Websites

TCP connect time towards websites under LTE.

∼ 90% of the time, www.facebook.com and www.youtube.com can be reached in less than 100 ms from a client’s device. for www.google.fi and www.elisa.net, only 80% and 76% of the TCP connection test are below 100 ms, respectively.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 17 / 21

slide-22
SLIDE 22

Data Analysis & Results

TCP Connect time — by destination ASN from LTE networks

Requests towards www.youtube.com served by the ISPs cache are faster than those served by Google CDN. www.facebook.com does not hit any caches in the ISP network —

slower TCP connect time than www.youtube.com and www.google.fi

Caching can improves the fetch time of small files

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 18 / 21

slide-23
SLIDE 23

Data Analysis & Results

TCP Connect time— by destination ASN from LTE networks

0.00 0.25 0.50 0.75 1.00 −100−80 −60 −40 −20 20 40 60 80 100 TCP connect time (ms)

∆= [AS719(Elisa) − AS15169(Google)]

LTE network

TCP connect time towards www.google.fi showing the latency difference between ISP cache - Elisa (AS719) and CDN - Google (AS15169) using LTE.

Values on the negative scale indicate that ISP cache is faster ∼ 70% of TCP connect time towards www.google.fi achieve lower latency when they hit ISP cache.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 19 / 21

slide-24
SLIDE 24

Conclusion

slide-25
SLIDE 25

Conclusion

Conclusion

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 20 / 21

slide-26
SLIDE 26

Conclusion

Conclusion

1

DNS lookup & TCP connect times varies for different websites, even when the same radio technology is accessed.

ISP caches & DNS server proximity

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 20 / 21

slide-27
SLIDE 27

Conclusion

Conclusion

1

DNS lookup & TCP connect times varies for different websites, even when the same radio technology is accessed.

ISP caches & DNS server proximity

2

Caches entries closer to the ISP could significantly improve TCP connect time.

the proximity of DNS server to the subscriber has a higher impact on DNS lookup time performance.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 20 / 21

slide-28
SLIDE 28

Conclusion

Conclusion

1

DNS lookup & TCP connect times varies for different websites, even when the same radio technology is accessed.

ISP caches & DNS server proximity

2

Caches entries closer to the ISP could significantly improve TCP connect time.

the proximity of DNS server to the subscriber has a higher impact on DNS lookup time performance.

3

Network radio technology:

LTE offers considerably low latency compared to legacy radio technologies.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 20 / 21

slide-29
SLIDE 29

Conclusion

Conclusion

1

DNS lookup & TCP connect times varies for different websites, even when the same radio technology is accessed.

ISP caches & DNS server proximity

2

Caches entries closer to the ISP could significantly improve TCP connect time.

the proximity of DNS server to the subscriber has a higher impact on DNS lookup time performance.

3

Network radio technology:

LTE offers considerably low latency compared to legacy radio technologies.

4

Packet loss using ping test can be underestimated

e.g., if # of packets to be sent per pingtest instance < 5. consider increasing the number of packets per ping test instance for better results.

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 20 / 21

slide-30
SLIDE 30

Conclusion

Conclusion

1

DNS lookup & TCP connect times varies for different websites, even when the same radio technology is accessed.

ISP caches & DNS server proximity

2

Caches entries closer to the ISP could significantly improve TCP connect time.

the proximity of DNS server to the subscriber has a higher impact on DNS lookup time performance.

3

Network radio technology:

LTE offers considerably low latency compared to legacy radio technologies.

4

Packet loss using ping test can be underestimated

e.g., if # of packets to be sent per pingtest instance < 5. consider increasing the number of packets per ping test instance for better results.

ermias.walelgne@aalto.fi

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 20 / 21

slide-31
SLIDE 31

Conclusion

References

References

Brownlee, N., & claffy, k. (2002). Internet stream size distributions. In Proceed- ings of the 2002 acm sigmetrics international conference on measurement and modeling of computer systems (pp. 282–283). New York, NY, USA:

  • ACM. Retrieved from http://doi.acm.org/10.1145/511334.511381

doi: 10.1145/511334.511381 Huang, J., Qian, F., Guo, Y., Zhou, Y., Xu, Q., Mao, Z. M., . . . Spatscheck, O. (2013). An In-depth Study of LTE: Effect of Network Protocol and Ap- plication Behavior on Performance. In (pp. 363–374). Retrieved from http://doi.acm.org/10.1145/2486001.2486006 doi: 10.1145/ 2486001.2486006 Ramachandran, S. (2010). Lets make the web faster. Google. http://code.

  • google. com/speed/articles/webmetrics. html.

Rula, J. P., & Bustamante, F. E. (2014). Behind the Curtain: The Importance of Replica Selection in Next Generation Cellular Networks. In (pp. 135–136).

Factors Affecting Performance of Web Flows in Cellular Networks IFIP Networking’18, Zurich 21 / 21