3 ECE Department Introduction Most popular user-generated video service. 800 million unique users, billions of hours of videos. Globally distributed levels of caches. Difficult to predict videos watched compared to other video services (E.g., Netflix or Hulu).
5 ECE Department Motivation (Contd..) Most people select their videos from related list. Out of the 20 related list offered, most people tend to select videos from the Top 10. Caching and Prefetching of related videos are shown to be effective. Streaming quality and network load reduction can be achieved.
6 ECE Department Related List Reordering Cache hit rate increases by 2 to 5 times by reordering.
7 ECE Department Objective To find out if the related videos list offered change based on region and time. How much of the related video list changes? What is the impact of these related video changes on caching or prefetching?
8 ECE Department Experiment Setup PlanetLab Measurement for global analysis. 4 different regions (US, EU, AS, SA). US – 197 nodes, EU- 243 nodes, AS – 62 nodes and SA – 17 nodes. 519 total nodes and 100 random videos.
9 ECE Department Metrics Content Change Order Change CC = 2, OC = 4 CC = 0, OC = 2
11 ECE Department Analysis Results (Content Change) US Region EU Region
12 ECE Department Order Change Results US Region EU Region
13 ECE Department Daily Change (Content Change) US Region EU Region
14 ECE Department Daily Changes (Order Change) US Region EU Region
15 ECE Department Impact on Caching? Duration3 days Total Requests105339 Related Videos47986
16 ECE Department Impact on Regional Differences 35% CC related list difference of at least 2 for Top 5 related videos. 60% requests for Top 5 videos. Leads to 21% additional caching of content. 65% related list difference of at least 2 for the bottom half. But only 20% requests for bottom half of related videos. For OC related list difference, 60% of at least 3 in order for Top 5. Affects the related list reordering technique. OC increases to 90% for bottom half of list.
17 ECE Department Impact on Client Differences 42% hit rate of client caching/prefetching for Top 10 related videos. Related list differences of at least 3 for Top 10 is about 20% and 40% for bottom half. Leads to 8% additional caching/prefetching of content at client. 6% improvement in cache hit rate for bottom half but 40% increase in list difference.
18 ECE Department Conclusion We perform a global study on related list behavior. We find that the list changes from region to region and also on the same client on daily basis. This list difference reduces the efficiency of caching on the edge or at client. By the analysis, we find that caching Top half of related list offers better trade-off between cache hit rate and list changes.
19 ECE Department Future Work How related list offered differs based on different factors? Such as popularity of videos, view count, region etc., What parts of related list transmitted to clients are already stored in YouTube Cache? We can use the cache reordering approach to know if it is delivered by cache.