Tracing the Path to YouTube - Introduction A Quantification of Path - PowerPoint PPT Presentation

Tracing the Path to YouTube - Introduction A Quantification of Path Lengths and Latencies towards Content Caches Motivation Research Questions Methodology Accepted for publication in IEEE Communications Magazine Analysis (Pre-print: http://in.tum.de/~doan/2018-yt-traces.pdf ) Paths Deltas Destination Pairs Content Caches Trinh Viet Doan, Ljubica Pajević, Conclusion Vaibhav Bajpai, Jörg Ott Chair of Connected Mobility Technical University of Munich RIPE 77, Amsterdam October 17, 2018 1 / 18

Introduction Motivation Research Questions Methodology Analysis Paths Deltas Destination Pairs Content Caches Introduction Conclusion 2 / 18

Introduction TCP Connect Times Motivation 0 ∆t (ms) Introduction Web Motivation −5 Research Questions Previous work [2]: Methodology TCP Connect Times 0.0 ∆t (ms) Analysis ◮ Measuring YouTube −0.1 Audio −0.2 Paths Video performance for popular −0.3 Deltas −0.4 Destination Pairs videos Content Caches Prebuffering Duration Conclusion ◮ Performance over IPv6 is ∆p (ms) 0 −40 worse than over IPv4 −80 −120 ◮ Speculation: Startup Delay 0 Content caches not ∆s (ms) −100 −200 dual-stacked? −300 −400 Jul Jul Jan Jan Jan Jul 2015 2016 2017 Figure 1: Difference of YouTube performance metrics over IPv4 and IPv6 3 / 18

Introduction Research Questions Introduction Motivation Research Questions Methodology Analysis Paths Deltas Destination Pairs 1. How far are content caches from users? Content Caches Conclusion 2. How much benefit do these caches provide? 3. How do these metrics compare quantitatively over IPv4 and IPv6? 4 / 18

Introduction Motivation Research Questions Methodology Analysis Paths Deltas Destination Pairs Content Caches Methodology Conclusion 5 / 18

Methodology Measurement Setup Introduction Motivation Research Questions Methodology Analysis Paths Deltas Destination Pairs Content Caches Conclusion Figure 3: Example of measurement probe: SamKnows Whitebox 8.0 1 Figure 2: Map of SamKnows probes ◮ ≈ 100 probes deployed around the world since 2014 ◮ Deployed in dual-stacked residential networks, NRENs, business networks, research labs, data centers, IXPs, ... ◮ Active measurement studies from fixed-line networks 1 https://blog.samknows.com/new-testing-superfast-broadband-27a7abcf1303 [accessed 2018-08-07] 6 / 18

Methodology Targets and Metrics Introduction Motivation Research Questions Methodology ◮ Hourly traceroute measurements over IPv4 & IPv6 Analysis Paths ◮ Using scamper [3] for paris traceroute over ICMP Deltas Destination Pairs Content Caches ◮ Targets: YouTube media servers Conclusion ◮ Media servers identified by youtube test [1] that mimics video streaming from YouTube ◮ DNS resolution for this streaming directly on the probe ⇒ Redirected to best/closest cache, determined by YouTube ◮ Identified IP addresses of media servers to scamper for measurements ◮ Time period: since May 2016 7 / 18

Introduction Motivation Research Questions Methodology Analysis Paths Deltas Destination Pairs Content Caches Analysis Conclusion 8 / 18

Analysis Paths 1.0 Introduction RTT (IPv6) [ms] Motivation 0.8 RTT (IPv4) [ms] Research Questions TTL (IPv6) Methodology 0.6 TTL (IPv4) CDF Analysis 0.4 Paths Deltas 0.2 Destination Pairs Content Caches 0.0 Conclusion 1 10 100 Figure 4: CDF of median IP path TTL and RTT ◮ Comparable number of paths observed ◮ 78% with TTL ≤ 12 (IPv4), ≤ 11 (IPv6) → IPv6 paths more often shorter ◮ 74% with RTT ≤ 25 ms (IPv4), 72% over (IPv6) → IPv6 more often slower 9 / 18

Analysis Deltas Introduction However, no direct comparison possible ⇒ look at destination pairs Motivation Research Questions Methodology unit_id dtime source destination status ttl endpoint rtt Analysis 239416 2016-06-07 16:45:35 2001:67c:_:_:_:_:fef0:d612 2a00:1450:400f:f::a COMPLETED 9 2a00:1450:400f:f::a 10.522 Paths 239416 2016-06-07 16:45:36 10.0.1.3 83.255.235.81 COMPLETED 7 83.255.235.81 13.178 Deltas Destination Pairs Content Caches Conclusion ∆ TTL = TTL IPv 4 − TTL IPv 6 ∆ RTT = RTT IPv 4 − RTT IPv 6 Figure 5: Example for a destination pair 10 / 18

Analysis Destination Pairs: General Introduction IPv6 slower IPv6 faster Motivation Research Questions 1.0 Methodology 0.8 Analysis Paths 0.6 CDF Deltas Destination Pairs 0.4 Content Caches 0.2 Conclusion RTT delta [ms] TTL delta 0.0 1000 100 10 1 0 1 10 100 1000 delta Figure 6: CDF of median destination pair deltas ◮ TTL: ◮ RTT: ◮ 27% with ∆ TTL < 0 ◮ ≈ 50% with ∆ RTT < 0 ◮ 33% with ∆ TTL = 0 ◮ ≈ 50% with ∆ RTT > 0 ◮ 40% with ∆ TTL > 0 11 / 18

Analysis Destination Pairs: General Introduction IPv6 slower IPv6 faster Motivation Research Questions 1.0 Methodology 0.8 Analysis Paths 0.6 CDF Deltas Destination Pairs 0.4 Content Caches 0.2 Conclusion RTT delta [ms] TTL delta 0.0 1000 100 10 1 0 1 10 100 1000 delta Figure 6: CDF of median destination pair deltas ◮ Overall: ◮ TTL: 91% within [-5; +5] ◮ RTT: 91% within [-20; +20] ms 11 / 18

Analysis Content Caches Introduction Motivation Research Questions Methodology ◮ Content caches usually deployed within ISP networks Analysis Paths ◮ In close proximity to users to reduce latency Deltas Destination Pairs Content Caches Conclusion 12 / 18

Analysis Content Caches Introduction Motivation Research Questions Methodology ◮ Content caches usually deployed within ISP networks Analysis Paths ◮ In close proximity to users to reduce latency Deltas Destination Pairs Content Caches Conclusion ◮ How to identify caches? ◮ Matching AS numbers for source and destination → src ASN == dst ASN ◮ Reverse DNS lookups of destination IP addresses to retrieve human-readable hostnames → keywords: cache or ggc ◮ Lookups using RIPEstat 2 2 https://stat.ripe.net/ 12 / 18

Analysis Destination Pairs: Caches Introduction Motivation Research Questions Methodology Possible scenarios for identification of caches when comparing between different Analysis Paths address families. Deltas Destination Pairs Content Caches IPv6 Conclusion Cache No Cache IPv4 both IPv4 only Cache O △ IPv6 only neither No Cache ♦ � 13 / 18

Analysis Destination Pairs: Caches Introduction Motivation IPv6 slower IPv6 faster IPv6 slower IPv6 faster Research Questions 1.0 1.0 Methodology 0.8 0.8 0.6 Analysis 0.6 CDF CDF Paths 0.4 IPv4 only IPv4 only 0.4 IPv6 only IPv6 only Deltas 0.2 0.2 Both versions Both versions Destination Pairs Neither version Neither version 0.0 0.0 Content Caches 15 10 5 0 5 10 15 1000 100 10 1 0 1 10 100 1000 Conclusion TTL delta RTT delta [ms] Figure 7: CDF of median destination pair deltas (split) 13 / 18

Analysis Destination Pairs: Caches Introduction Motivation IPv6 slower IPv6 faster IPv6 slower IPv6 faster Research Questions 1.0 1.0 Methodology 0.8 0.8 0.6 Analysis 0.6 CDF CDF Paths 0.4 IPv4 only IPv4 only 0.4 IPv6 only IPv6 only Deltas 0.2 0.2 Both versions Both versions Destination Pairs Neither version Neither version 0.0 0.0 Content Caches 15 10 5 0 5 10 15 1000 100 10 1 0 1 10 100 1000 Conclusion TTL delta RTT delta [ms] Figure 7: CDF of median destination pair deltas (split) ◮ IPv4 cache only ( △ ): shifted to left side; RTT lower over IPv4 for ≈ 80% 13 / 18

Analysis Destination Pairs: Caches Introduction Motivation IPv6 slower IPv6 faster IPv6 slower IPv6 faster Research Questions 1.0 1.0 Methodology 0.8 0.8 0.6 Analysis 0.6 CDF CDF Paths 0.4 IPv4 only IPv4 only 0.4 IPv6 only IPv6 only Deltas 0.2 0.2 Both versions Both versions Destination Pairs Neither version Neither version 0.0 0.0 Content Caches 15 10 5 0 5 10 15 1000 100 10 1 0 1 10 100 1000 Conclusion TTL delta RTT delta [ms] Figure 7: CDF of median destination pair deltas (split) ◮ IPv4 cache only ( △ ): shifted to left side; RTT lower over IPv4 for ≈ 80% ◮ IPv6 cache only ( � ): paths shorter to IPv6 caches compared to IPv4 no-cache destinations; yet still higher RTTs in most cases 13 / 18

Analysis Destination Pairs: Caches Introduction Motivation IPv6 slower IPv6 faster IPv6 slower IPv6 faster Research Questions 1.0 1.0 Methodology 0.8 0.8 0.6 Analysis 0.6 CDF CDF Paths 0.4 IPv4 only IPv4 only 0.4 IPv6 only IPv6 only Deltas 0.2 0.2 Both versions Both versions Destination Pairs Neither version Neither version 0.0 0.0 Content Caches 15 10 5 0 5 10 15 1000 100 10 1 0 1 10 100 1000 Conclusion TTL delta RTT delta [ms] Figure 7: CDF of median destination pair deltas (split) ◮ IPv4 cache only ( △ ): shifted to left side; RTT lower over IPv4 for ≈ 80% ◮ IPv6 cache only ( � ): paths shorter to IPv6 caches compared to IPv4 no-cache destinations; yet still higher RTTs in most cases ◮ Both ( O ): deltas converging towards zero; 60% of the time faster over IPv4, 40% of the time faster over IPv6, however ≈ 80% within [-1,+1] ms 13 / 18