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 University of Munich (TUM), † SamKnows IEEE INFOCOM, July 6–9, 2020 1 / 16
Motivation Methodology Analysis Content Delivery Content Caches Conclusions Motivation 2 / 16
Motivation Motivation Methodology ◮ Video streaming accounts for majority of Internet traffic Analysis ◮ YouTube 8.7%, Netflix 12.6% of all global downstream traffic 1 as of 2019 Content Delivery Content Caches ◮ Introduction of Netflix Open Connect in 2011 Conclusions ◮ 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 connectivity 2 as of 2020 ⇒ Longitudinal measurement study on Netflix OCAs from multiple vantage points with focus on content delivery over IPv6 and content caches 1 Sandvine, Global Internet Phenomena Report 2019, https://bit.ly/3cvN5Qi 2 Google IPv6 Statistics, https://www.google.com/intl/en/ipv6/statistics.html 3 / 16
Motivation Methodology Analysis Content Delivery Content Caches Conclusions Methodology 4 / 16
Methodology Motivation Methodology Analysis Content Delivery Content Caches Conclusions Source: https://samknows.one/hc/en-gb/articles/360000451757 ◮ Roughly 100 probes deployed in 74 origin ASes (primarily home ◮ Hardware: networks) dedicated SamKnows Whiteboxes ◮ Dual-stack with native IPv6 (formerly TP-Link routers with connectivity, i.e., no transition customized OpenWrt) mechanisms at access network ◮ Running set of measurement tests ◮ Measurement period: periodically July 2016 – April 2019 5 / 16
Methodology Motivation Tests: Methodology ◮ netflix test and traceroute ( scamper ) to Netflix OCAs Analysis Content Delivery ◮ speedtest to Measurement Lab (M-Lab) servers Content Caches Conclusions ◮ Schedule: tests run every hour over IPv4 and IPv6 each (“pair”) netflix test: Control Plane 1. Authenticate with control plane api.fast.com netflix.com 2. Receive list of OCAs determined JSON manifest by fast.com API OCA1_URL OCA1_URL Netflix OCA … 3. Connect to OCA SamKnows 4. (Repeatedly) download 25 MB Probe file for 20 sec to simulate streaming 6 / 16
Motivation Methodology Analysis Content Delivery Analysis: Content Delivery Content Caches Conclusions ◮ How does Netflix content delivery perform w.r.t. latency and throughput over both address families? ◮ Do users benefit or suffer from downloading Netflix content over IPv6 compared to IPv4? 7 / 16
IPv6 Preference IPv6 Preference Motivation 100.0% 99.5% Methodology 99.0% Analysis 98.5% Content Delivery Content Caches 0 1 2 3 4 5 6 7 8 9 1011121314151617181920212223 Conclusions Hours of the Day ◮ 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
TCP Connect Time TCP Connect Time IPv6 slower IPv6 faster Motivation 1.0 0.8 Methodology 2016 0.6 CDF 2017 Analysis 0.4 2018 Content Delivery 0.2 2019 Content Caches 0.0 Conclusions 10 2 10 1 10 0 0 10 0 10 1 10 2 Delta (ms) Prebuffering Duration 1.0 ◮ Median TCP connect time deltas ( tconn IPv 4 − tconn IPv 6 ) of probes per day 0.8 2016 0.6 CDF 2017 ◮ Most differences within ± 10 ms, converging toward 0 ms 0.4 2018 0.2 2019 0.0 ◮ Across all years, 45%–48% of samples faster over IPv6 10 0 0 10 0 10 3 10 2 10 1 10 1 10 2 10 3 ◮ Absolute connect times improved by 40% over the years (25 ms → 15 ms) Delta (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
Throughput Throughput [Netflix OCA] IPv6 higher throughput IPv6 lower throughput ◮ Median throughput deltas Motivation 1.0 ( tp IPv 4 − tp IPv 6 ) of 0.8 Methodology probes per day 0.6 CDF Analysis 2016 0.4 2017 Content Delivery ◮ IPv6 higher throughput 0.2 2018 Content Caches 2019 0.0 Conclusions than IPv4 in only 10 1 10 0 0 10 0 10 1 17%–25% of samples Delta (MB/s) ◮ 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
Motivation Methodology Analysis Analysis: Content Caches Content Delivery Content Caches Conclusions ◮ 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
Content Caches TTL 1.0 Motivation ◮ Comparison of ISP caches 0.8 Methodology and CDN servers IPv4 Cache 0.6 CDF IPv4 CDN Analysis 0.4 IPv6 Cache ◮ Measurement distribution Content Delivery 0.2 IPv6 CDN Content Caches 0.0 similar over IPv4 and IPv6 Conclusions 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 ◮ 90% of ISP caches within TCP Connect Time 6 IP hops & 21 ms 1.0 IPv4 Cache 0.8 IPv4 CDN ◮ 90% of CDN servers within 0.6 CDF IPv6 Cache 0.4 IPv6 CDN 10–12 IP hops & 56–58 ms 0.2 0.0 ◮ Higher throughput with 0 10 0 10 1 10 2 caches: 10–11 → 32–34 MB/s ms 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
Content Caches Motivation ◮ TCP connect time deltas Methodology ( tconn CDN − tconn cache ), Analysis Content Delivery probe medians per day, Content Caches split by address family Conclusions ◮ 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
Motivation Methodology Analysis Content Delivery Content Caches Conclusions Conclusions 14 / 16
Conclusions: Limitations Motivation Methodology Analysis Content Delivery Content Caches Conclusions ◮ 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
Conclusions: Summary ◮ Latency and IP path lengths similar between both address families Motivation ⇒ High IPv6 preference, however, slight drops during peak hours Methodology Analysis ◮ Throughput increases over the years, though not specific to Netflix Content Delivery ◮ Over IPv6 lower than over IPv4 in most cases (>75%) Content Caches ◮ However, most deltas only within ± 1 MB/s Conclusions ◮ 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 Vaibhav Bajpai Trinh Viet Doan Sam Crawford bajpaiv@in.tum.de doan@in.tum.de sam@samknows.com 16 / 16
Recommend
More recommend
