20
Design and Analysis of Advanced Replacement Policies for WWW Caching Kai Cheng, Yusuke Yokota, Yahiko Kambayashi Department of Social Informatics Graduate School of Informatics, Kyoto University

Design and Analysis of Advanced Replacement Policies for WWW Caching Kai Cheng, Yusuke Yokota, Yahiko Kambayashi Department of Social Informatics Graduate

Embed Size (px)

Citation preview

Design and Analysis of Advanced  Replacement   Policies for WWW

Caching

Kai Cheng, Yusuke Yokota, Yahiko Kambayashi

Department of Social Informatics

Graduate School of Informatics, Kyoto University

3 March 2000 DEWS 2000, Biwako 2

Contents

• Introduction• Hybrid Replacement Policies for Web Caching • Cache Structure for Hybridizing Various Policies• PSS-W: Hybridizing LRU, LFU and SIZE • Experimental Evaluation and Analysis• Conclusion and Future Work

Conceptual and Methodological Support for ADVACED POLICIES managing Large Cache Space, Object Space, implementing Sophisticated Control Logic, with high performance and low overhead

3 March 2000 DEWS 2000, Biwako 3

Introduction (1): Caching on the World Wide Web

• World Wide Web or World Wide Wait– Popular: 75% of Internet Backbone Traffic [1]– Costly: Repeated Transfer of Same Content

• Caching the Frequently Used Close to Users – Sharing Content Over Time by a Given User – Sharing Content Between Many Users

• Benefits from Web Caching – Improved Response Times– Reduced Network Load– Increased Server Throughput

World Wide Wait

3 March 2000 DEWS 2000, Biwako 4

Introduction (2) What to Cache : Replacement Policies• Cache All ()

– Unrealistic: Persistent Cache -- New Contents Increase Ceaselessly

– Unnecessary: Reference Locality -- Only A Small Part of Sites or Documents Popular

• Choose to Cache ()– Bring in Highly Potential Objects– Displace Less Potential Ones

• Replacement Policies– Objectives: Hit Rate, Byte Hit Rate, etc– Bases: Reference Locality; Size, Cost, etc

3 March 2000 DEWS 2000, Biwako 5

Hybrid Replacement Policies for Web Caching

• Why Hybrid Caching for Web– Manage Large Cache Space, Object Space– Implement Sophisticated Control Logic– Avoid “Cache Pollution”– Reduce “Early Eviction”

• Hybridize Several Policies (Examples)– Segmented LRU [2]: Hybridize LFU and LRU– Size-Adjusted LRU [3]: Hybridize SIZE and LRU – Segmented FIFO [4] : Hybridize LFU and FIFO

3 March 2000 DEWS 2000, Biwako 6

Cache Structure for Hybridizing Various Policies (1) Basic

• Cache: A Definition – Cache Space

– Object Space

– Replacement Policy

– Constraints : admission, freshness, etc.

• Cache Structure– Monolithic: monolithic cache space, single object list

– Distributed : segmented cache space, multiple objects lists

S

O

CP

3 March 2000 DEWS 2000, Biwako 7

Cache Structure for Hybridizing Various Policies (2) Our Proposal

• Classification Rules – Clustering Web Objects based on sizes, types, usages

– Localizing Replacement Decisions

• Unit Caches– Maintaining Similar Web Objects

• Central Cache– Solving Non-Determinism of Replacement Results

– Controlling “Cache Pollution”

3 March 2000 DEWS 2000, Biwako 8

Cache Structure for Hybridizing Various Policies (3)

Object Space

Cache Space

ClassificationRule

Unit Cache 0

Unit Cache 1

Unit Cache 2

Central Cache

Evictions

Hits

3 March 2000 DEWS 2000, Biwako 9

PSS-W: Hybridizing LRU, LFU and SIZE

• Size-Adjusted LRU : SIZE plus LRU– Largest First– Least Recently Used First

• Segmented LRU : LFU + LRU– Least Frequently Used First – Least Recently Used First

• PSS-W– Least Frequently Used First – Least Recently Used First– Largest First

Max {Atime x Size}

Max {Atime x Size / Nref}

LRU

LFU

SIZ

E

3 March 2000 DEWS 2000, Biwako 10

Pyramidal Selection Scheme (PSS) For Size-Adjusted LRU

Classification Rules Based on Size OnlyClassification Rules Based on Size Only

Objects Fixed in a Single LRU queueObjects Fixed in a Single LRU queue

3 March 2000 DEWS 2000, Biwako 11

Fig.1.Object With Uneven Accesses Being Deleted in PSS

Popular Objects Aged Out in PSS

3 March 2000 DEWS 2000, Biwako 12

PSS with aWarded Sizes (PSS-W)

Classification Rules Based on Size and FrequencyClassification Rules Based on Size and Frequency

Max {Size x Atime}

Objects Moving among LRU queuesObjects Moving among LRU queues

3 March 2000 DEWS 2000, Biwako 13

Fig. 2. Objects With Uneven Accesses Stay in PSS-W

Popular Object Stay Long in PSS-W

3 March 2000 DEWS 2000, Biwako 14

4. Trace-Driven Simulations

Fig. 3. Objects With More Accesses Tend More Popular

3 March 2000 DEWS 2000, Biwako 15

Fig.4. Significant Improvement in Byte Hit Rate

Bye Hit Rate Improvement to PSS

3 March 2000 DEWS 2000, Biwako 16Fig.4. No Harm To Hit Rate

Better Hit Rate than PSS

3 March 2000 DEWS 2000, Biwako 17

Fig.6. Hit Rate Champion

Outstanding Hit Rate

3 March 2000 DEWS 2000, Biwako 18

Fig.7. Competitive Byte Hit Rate

Competitive Byte Hit Rate

3 March 2000 DEWS 2000, Biwako 19

Conclusion and Future Work

• Highlighted Importance of Hybrid Policies for Web Caching

• Proposed a Cache Structure to Facilitate Hybridizing Various Policies

• Developed PSS-W : Hybridizing SIZE, LRU and LFU

• Extending for Application-Level Metrics

3 March 2000 DEWS 2000, Biwako 20

References

① K. Tompson et al. Wide-Area Internet Pattern and Characteristics, IEEE Network, 11(6), 1997

② R. Karedla, et al. Caching Strategies for Improving Disk System Performance, IEEE Computer, 27(3), 1994

③ C. Aggarwal et al. Caching on the World Wide Web, IEEE KDE 11(1), 1999

④ R. Tuner et al. Segmented FIFO Page Replacement. Proc. ACM SIGMETRICS’81 pp. 48-51 1981