A Longitudinal View of Netflix: Content Delivery over IPv6 - - PowerPoint PPT Presentation

▶

May 27, 2023 296 likes •471 views

A Longitudinal View of Netflix: Content Delivery over IPv6 Motivation Methodology and Content Cache Deployments Analysis Content Delivery Content Caches Conclusions Trinh Viet Doan * , Vaibhav Bajpai * , Sam Crawford * Technical

SLIDE 1

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

A Longitudinal View of Netflix: Content Delivery over IPv6 and Content Cache Deployments

Trinh Viet Doan, Vaibhav Bajpai, Sam Crawford†

*Technical University of Munich (TUM), †SamKnows

IEEE INFOCOM, July 6–9, 2020

1 / 16

SLIDE 2

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Motivation

2 / 16

SLIDE 3

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Motivation

◮ Video streaming accounts for majority of Internet traffic

◮ YouTube 8.7%, Netflix 12.6% of all global downstream traffic1 as of 2019

◮ Introduction of Netflix Open Connect in 2011 ◮ Open Connect Appliances (OCAs) deployed within IXPs and ISPs to bring

content to the edge

◮ YouTube and Netflix as main drivers of IPv6

◮ 31.5% of Google users with available IPv6 connectivity2 as of 2020

⇒ Longitudinal measurement study on Netflix OCAs from multiple vantage points with focus on content delivery over IPv6 and content caches

1Sandvine, Global Internet Phenomena Report 2019, https://bit.ly/3cvN5Qi 2Google IPv6 Statistics, https://www.google.com/intl/en/ipv6/statistics.html 3 / 16

SLIDE 4

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Methodology

4 / 16

SLIDE 5

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Methodology

◮ Roughly 100 probes deployed in

74 origin ASes (primarily home networks)

◮ Dual-stack with native IPv6

connectivity, i.e., no transition mechanisms at access network

◮ Measurement period:

July 2016 – April 2019

Source: https://samknows.one/hc/en-gb/articles/360000451757

◮ Hardware:

dedicated SamKnows Whiteboxes (formerly TP-Link routers with customized OpenWrt)

◮ Running set of measurement tests

periodically

5 / 16

SLIDE 6

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Methodology

Tests:

◮ netflix test and traceroute (scamper) to Netflix OCAs ◮ speedtest to Measurement Lab (M-Lab) servers ◮ Schedule: tests run every hour over IPv4 and IPv6 each (“pair”)

netflix test:

1. Authenticate with control plane
2. Receive list of OCAs determined

by fast.com API

3. Connect to OCA
4. (Repeatedly) download 25 MB

file for 20 sec to simulate streaming

SamKnows Probe

JSON manifest OCA1_URL OCA1_URL …

Netflix OCA netflix.com api.fast.com Control Plane

6 / 16

SLIDE 7

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Analysis: Content Delivery

◮ How does Netflix content delivery perform w.r.t. latency and throughput

ver both address families?

◮ Do users benefit or suffer from downloading Netflix content over IPv6

compared to IPv4?

7 / 16

SLIDE 8

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

IPv6 Preference

0 1 2 3 4 5 6 7 8 9 1011121314151617181920212223 Hours of the Day 98.5% 99.0% 99.5% 100.0% IPv6 Preference ◮ Calculate IPv6 preference based on Happy Eyeballs [RFC 8305] ◮ Group measurement pairs by local time ◮ Overall, very high IPv6 preference, i.e., IPv6 slower by at most 250 ms ◮ However, IPv6 connections preferred less during peak hours (18:xx–23:xx)

IPv6 preferred in nearly all cases when connecting to Netflix OCAs, although preference slightly drops during peak hours.

8 / 16

SLIDE 9

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

TCP Connect Time

102 101 100 100 101 102 0.0 0.2 0.4 0.6 0.8 1.0 CDF IPv6 slower IPv6 faster TCP Connect Time

2016 2017 2018 2019

103 102 101 100 0 100 101 102 103 Delta (ms) 0.0 0.2 0.4 0.6 0.8 1.0 CDF Prebuffering Duration

2016 2017 2018 2019

Delta (ms)

◮ Median TCP connect time deltas (tconnIPv4 − tconnIPv6) of probes per day ◮ Most differences within ±10 ms, converging toward 0 ms ◮ Across all years, 45%–48% of samples faster over IPv6 ◮ Absolute connect times improved by 40% over the years (25 ms → 15 ms)

TCP connect times over IPv4 and IPv6 converge to similar values as of 2019, with roughly 50% of the samples being faster/slower over IPv6.

9 / 16

SLIDE 10

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Throughput

◮ Median throughput deltas

(tpIPv4 − tpIPv6) of probes per day

◮ IPv6 higher throughput

than IPv4 in only 17%–25% of samples

101 100 100 101 Delta (MB/s) 0.0 0.2 0.4 0.6 0.8 1.0 CDF IPv6 higher throughput IPv6 lower throughput Throughput [Netflix OCA]

2016 2017 2018 2019

◮ However, 70–75% of samples within ±1 MB/s ◮ Achieved throughput increases over the years, though not specific to Netflix

◮ Similar observations for speedtest measurements toward M-Lab servers

Although most of the deltas are within ±1 MB/s, IPv6 shows lower throughput than IPv4 in 75–83% of the samples (not specific to Netflix).

10 / 16

SLIDE 11

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Analysis: Content Caches

◮ How do IP path lengths and latency compare between ISP caches and

deployments outside the ISP boundary?

◮ How do path lengths and latency differ over IPv4 and IPv6?

◮ How do content caches at the edge benefit content delivery?

◮ How do these benefits compare over IPv4 and IPv6? 11 / 16

SLIDE 12

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Content Caches

◮ Comparison of ISP caches

and CDN servers

◮ Measurement distribution

similar over IPv4 and IPv6

◮ 90% of ISP caches within

6 IP hops & 21 ms

◮ 90% of CDN servers within

10–12 IP hops & 56–58 ms

◮ Higher throughput with

caches: 10–11 → 32–34 MB/s

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 0.0 0.2 0.4 0.6 0.8 1.0 CDF TTL

IPv4 Cache IPv4 CDN IPv6 Cache IPv6 CDN

100 101 102 ms 0.0 0.2 0.4 0.6 0.8 1.0 CDF TCP Connect Time

IPv4 Cache IPv4 CDN IPv6 Cache IPv6 CDN

ISP caches are reachable within 6 IP hops and roughly 20 ms; IP path lengths are shorter by 40–50%, TCP connect times lower by 64%, throughput higher by factor of three.

12 / 16

SLIDE 13

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Content Caches

◮ TCP connect time deltas

(tconnCDN − tconncache), probe medians per day, split by address family

◮ Cache benefits larger over

IPv4 than over IPv6

◮ Longitudinally: benefits of caches grow over IPv4, however, cache benefits

become smaller over IPv6

◮ Latency similar when caches dual-stacked; however, benefits more

pronounced for IPv4-only caches compared to IPv6-only caches Latency benefits toward ISP caches have become more pronounced over IPv4 and less pronounced over IPv6 through the years.

13 / 16

SLIDE 14

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

14 / 16

SLIDE 15

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Conclusions: Limitations

◮ Caches identified by matching ASNs, does not account for peering ◮ Location of probes biased toward EU and NA

◮ Limits analysis of regional differences 15 / 16

SLIDE 16

Motivation Methodology Analysis

Content Delivery Content Caches

Conclusions

Conclusions: Summary

◮ Latency and IP path lengths similar between both address families

⇒ High IPv6 preference, however, slight drops during peak hours

◮ Throughput increases over the years, though not specific to Netflix

◮ Over IPv6 lower than over IPv4 in most cases (>75%) ◮ However, most deltas only within ±1 MB/s

◮ Caches reachable within 6 IP hops and 20 ms

◮ IP path lengths shorter by 40–50% ◮ Latency lower by 64% ◮ Throughput higher by factor of three ◮ Latency benefits more pronounced over IPv4 compared with IPv6

Dataset and analysis scripts online:

https://github.com/tv-doan/infocom-2020-netflix Trinh Viet Doan doan@in.tum.de Vaibhav Bajpai bajpaiv@in.tum.de Sam Crawford sam@samknows.com

16 / 16

A Longitudinal View of Netflix: Content Delivery over IPv6 and Content Cache Deployments

Trinh Viet Doan*, Vaibhav Bajpai*, Sam Crawford†

IEEE INFOCOM, July 6–9, 2020

Motivation

Motivation

content to the edge

⇒ Longitudinal measurement study on Netflix OCAs from multiple vantage points with focus on content delivery over IPv6 and content caches

Methodology

Methodology

74 origin ASes (primarily home networks)

connectivity, i.e., no transition mechanisms at access network

July 2016 – April 2019

dedicated SamKnows Whiteboxes (formerly TP-Link routers with customized OpenWrt)

periodically

Methodology

Tests:

netflix test:

by fast.com API

file for 20 sec to simulate streaming

Analysis: Content Delivery

compared to IPv4?

IPv6 Preference

IPv6 preferred in nearly all cases when connecting to Netflix OCAs, although preference slightly drops during peak hours.

TCP Connect Time

TCP connect times over IPv4 and IPv6 converge to similar values as of 2019, with roughly 50% of the samples being faster/slower over IPv6.

Throughput

(tpIPv4 − tpIPv6) of probes per day

than IPv4 in only 17%–25% of samples

Although most of the deltas are within ±1 MB/s, IPv6 shows lower throughput than IPv4 in 75–83% of the samples (not specific to Netflix).

Analysis: Content Caches

deployments outside the ISP boundary?

Content Caches

ISP caches are reachable within 6 IP hops and roughly 20 ms; IP path lengths are shorter by 40–50%, TCP connect times lower by 64%, throughput higher by factor of three.

Content Caches

(tconnCDN − tconncache), probe medians per day, split by address family

IPv4 than over IPv6

become smaller over IPv6

pronounced for IPv4-only caches compared to IPv6-only caches Latency benefits toward ISP caches have become more pronounced over IPv4 and less pronounced over IPv6 through the years.

Conclusions

Conclusions: Limitations

Conclusions: Summary

⇒ High IPv6 preference, however, slight drops during peak hours

Dataset and analysis scripts online:

Trinh Viet Doan, Vaibhav Bajpai, Sam Crawford†