Categoryaware Hierarchical Caching for VideoonDemand Content on You

Motivation Content Delivery Network (CDN) cache Client

Motivation • Different content categories of video streams may have different cache demand •

Problem The problems that Adaptive Category-aware Designed Caching (ACDC) tries to solve are: 1.

Background – Cache Admission Policies • While the served content can be stored at

Background – Cache Eviction Policies • Cache eviction policy decides which item to be

System Design - Overview • Step 1. Content-awareness • Step 2. Distinct Caching Policies

System Design - Content-awareness 1. A video division enters the probationary cache - retrieves

System Design - Distinct Caching Policies per Category ACDC allows assigning individual cache policies

System Design - Cache-size Adaptation *when the IDs in ghost lists exceed the specified

System Design - Cache-size Adaptation • Division Size Adaptation Strategy 1. Smallest ghost list

Evaluation • Dataset • 10 million video requests to You. Tube videos • Cache

Evaluation - Best Caching Strategy per Category

Evaluation - ACDC Performance ACDC increases the CHR between 13% and 18% with a

Evaluation - Cache Layer Performance Gain = (CHR of ACDC / CHR of ARC)

Evaluation - Cache division size variation the storage size is increased from 1 TB

Evaluation - Eviction Policies • In many cases, ACDC can reach the best CHR

Conclusion Pros: • The idea of dynamically allocating cache space to different categories of

Slides: 19

Download presentation

Category-aware Hierarchical Caching for Video-on-Demand Content on You. Tube Christian Koch, Johannes Pfannmu ller, Amr Rizk, David Hausheer, Ralf Steinmetz Presenter: Tianshi Wang

Motivation Content Delivery Network (CDN) cache Client

Motivation • Different content categories of video streams may have different cache demand • E. g music being popular over weeks, while news being mostly popular on the same day This paper introduces Adaptive Content-Aware Designed Cache (ACDC)

Problem The problems that Adaptive Category-aware Designed Caching (ACDC) tries to solve are: 1. Differentiate the categories of video streaming (e. g. music, comedy, news …) 2. Adapt the cache storage share assigned per category under changing workload compositions cache Classifier Video File Music Entertainment People Comedy …

Background – Cache Admission Policies • While the served content can be stored at any of the traversed caches, the admission policy defines on which caches content is stored 1. Leave Copy Everywhere (LCE) 2. Leave a Copy Down (LCD) 3. Move Copy Down (MCD) 4. Probability Admission (Prob) 5. NHIT

Background – Cache Eviction Policies • Cache eviction policy decides which item to be evicted from the cache when the cache storage if full 1. LFU 1. A requested item is stored in the probationary part but its 2. LRU ID is already present in its ghost list – increase 3. SLRU Ghost List probationary cache size, decrease protected cache 4. ARC 1 1 probationary cache 1 protected cache size 2. An item is moved from the probationary to the protected cache and is present in the protected cache’s ghost list, the protected cache is increased and its probationary cache decreased

System Design - Overview • Step 1. Content-awareness • Step 2. Distinct Caching Policies per Category • Step 3. Cache-size Adaptation Cache Eviction Policy Classifier Video File Step 1 Music Entertainment People Comedy … Cache Size LRU TTL LFUDA Step 2 Step 3

System Design - Content-awareness 1. A video division enters the probationary cache - retrieves the video’s category ID (Youtube API) 2. If a video division in the probationary cache causes a cache hit - it is moved into a category-specific cache division Music (42. 5%) Entertainment (10. 1%) People & Blogs (8. 7%) Comedy (7. 8%) Miscellaneous (30. 8%)

System Design - Distinct Caching Policies per Category ACDC allows assigning individual cache policies on different categories

System Design - Cache-size Adaptation *when the IDs in ghost lists exceed the specified time-to-live (TTL) threshold, they are removed

System Design - Cache-size Adaptation • Division Size Adaptation Strategy 1. Smallest ghost list (SGL) 2. Largest ghost list (LGL) 3. Relatively smallest ghost list(RSGL) 4. Relatively largest ghost list (RLGL)

Evaluation • Dataset • 10 million video requests to You. Tube videos • Cache Hierarchies • Evaluation Metrics • Cache hit rate • Startup Delay • Write Operations

Evaluation - Best Caching Strategy per Category

Evaluation - ACDC Performance ACDC increases the CHR between 13% and 18% with a hierarchydependent optimum cache size of either 10 GB (H 1), 100 GB (H 2, H 3), or 1 TB (H 4)

Evaluation - Cache Layer Performance Gain = (CHR of ACDC / CHR of ARC) - 1 ACDC and ARC perform similarly during the morning hours. Overall, ACDC outperforms ARC which is especially visible at the peak-hours at noon

Evaluation - Cache division size variation the storage size is increased from 1 TB to 10 TB Shrink over night as many video segments exceed the maximum TTL value ACDC’s flexible storage allocation is leveraged

Evaluation - Eviction Policies • In many cases, ACDC can reach the best CHR • Improvement seems to be slight

Conclusion Pros: • The idea of dynamically allocating cache space to different categories of video stream is quite novel • The experiments in this paper reveal many interesting patterns of how the popularity of different categories of Youtube video streams change through the time Cons: • ACDC has slightly improvement compared to other traditional caching methods in many scenarios. • The methods including caching eviction policy and division size adaptation strategy are heuristic. There are no theoreticalanalysis on them. • Frequent Youtube API calls are required to classify categories. This could be an overhead for the servers.

System Design - Cache-size Adaptation