On the Intrinsic Locality Properties of Web Reference Streams
Download as PPTX, PDF1 like5,030 views
This document studies the intrinsic temporal locality properties of web request streams. It introduces metrics like entropy and coefficient of variation (CV) to measure popularity and correlation locality over time. The document shows how transformations like filtering, aggregation, and disaggregation impact these locality properties at different points in the web topology. Specifically, it finds that filtering preserves popularity locality but reduces correlation locality, while aggregation and disaggregation have varying effects on both metrics of temporal locality.
On the Intrinsic Locality Properties of Web Reference Streams
- 1. Intrinsic Temporal Locality Properties of Web Request Streams Rodrigo Fonseca (UC Berkeley) Virgílio Almeida (UFMG) Mark Crovella (Boston University) Bruno Abrahao (UFMG) IEEE INFOCOM 2003
- 2. Research Questions • Network measurements in highly dynamic environments – Understand the behavior and find intrinsic properties – Online algorithms for driving self- organization of system agents – Algorithms for dynamic adapting environments with high degree of uncertainty 2
- 3. Question • Study temporal locality at different points of the Web – Results: Framework + metrics: • New insights on the causes of temporal locality across different points of the Web topology 3
- 4. Web Measurements • Why? – Understand and improve current systems – Design new systems • Challenges – Complexity – Rate of change • Approach – Look for regularities, underlying principles – Abstractions 4
- 5. Previous Studies • Many advancements – Caching hierarchies – Replacement Policies for caches – Load distribution and balancing • However... – Focus on individual components – No structured view of the entire system 5
- 6. Stream centric view of the Web • Focus on request streams – Look for intrinsic properties – How they are altered – How their properties change in different points – How their intrinsic properties can influence the designs of components and of collections of components 6
- 7. Tranformations on Request Streams • Three types of transformations – Aggregation • Multiple sources – Disaggregation D A F • Multiple destinations – Filtering • Resulting stream = subset of input stream 7
- 8. Transformations Abstraction • Components may be abstracted by combinations of transformations Clients Proxy Cache Server D D F F A A 8
- 9. Approach • For a given intrinsic property of streams – Study the effects of the transformations on the property – Combine the effects to understand • Effects of components • Effects of collections of components • Properties at different points of the topology 9
- 10. Temporal Locality • Need of intrinsic metrics for the streams – We use virtual time • Two sources: – Popularity ...XAXBXCXDXE... • Preserved with reordering – Correlation ...AABBCCDDEE... • Lost with reordering 10
- 11. Measuring Popularity: Entropy – Traditionally Zipf’s Law: pi i- – We measure the deviation from Uniform popularity distribution – Entropy: N H pi log2 pi i 1 • pi determined empirically from frequencies • N is the number of distinct objects • If uniform, tends to log2(N) • If one object only is accessed, tends to 0 – Not the Entropy rate of the source 11
- 12. Entropy and Zipf’s Law 12
- 13. Entropy and Zip’f Law 13
- 14. Measuring Correlation: CV • Look at IAT distribution per object – Number of references between 2 references to the same object • Behavior – Correlation tendency to shorter IATs – No Correlation tends to geometric distribution • We measure the deviation from the geometric distribution for IAT – Coefficient of Variation: σ μ 14
- 15. IAT Distribution Scrambled Original 15
- 16. Metrics Summary • Entropy – High concentration: Low Entropy – Uniform distribution: High Entropy • CV – High Correlation: High CV – No correlation: CV ~ 1 16
- 17. Entropy vs. Hit Ratio • LRU cache simulation, size 5% 17
- 18. CV versus HR Difference • Difference between scrambled and original streams 18
- 19. Putting it all together • Effects of the transformations on the components of temporal locality 19
- 20. Effects of Filtering: Popularity 20
- 21. Effects of Filtering: Correlation 21
- 22. Effects of Aggregation and Disaggregation CV Entropy 22
- 23. Temporal Locality at Different Points of the topology Clients Proxies Servers 23
- 24. Effects of 24
- 25. Conclusions • Understanding of behavior • Transformations Framework allows for structured study and understanding of Web workloads characteristics 25