1. A Hybrid Caching Strategy for Streaming Media Files Jussara M. Almeida Derek L. Eager Mary K. Vernon University of Wisconsin-Madison University of Saskatchewan November 2001

2. Outline Characteristics of Streaming Media (SM) files Delivery of SM files Hypothesis and Assumptions Previous Caching Policies New Policy Performance Comparison New Caching Policies Conclusions and Future Work

3. Characteristics of SM Files Large file size –cache on disk Sustained I/O bandwidth –inserting and reading new content Clients access partial files –initial portion –favored segment –base + variable number of layers of layered encoding

4. Delivery of SM Files Unicast streaming: –server bandwidth is linear in client request rate –goal: maximize byte hit ratio Multicast streaming –save bandwidth –cost sharing introduces new tradeoffs

5. example: 10 distributed proxy servers each serving a local region, 100 requests (on avg) arrive per region during a given popular video need 7 streams per region, or 12 streams at the remote server Caching for Multicast Streams: Tradeoffs

6. caching popular content reduces the load on the remote server and network delivering popular content from the remote server amortizes the cost of a stream over more clients earlier portions of a popular video require more bandwidth and have less cost-sharing than later portions

7. New Caching Policies Research Hypothesis: popularity-based strategy will outperform replacement-based strategy –significant fraction of requests to uncached files may be for files that are accessed very sporadically Assumptions: –limited disk space implies limited disk bandwidth –proxy bandwidth for delivering cached streams is equal to min of proxy disk bw and proxy network bw (call this proxy disk bandwidth)

8. Current Web Caching Policies Replacement based (cache on each miss) Top replacement candidate is an ad-hoc combination of: –large files –least recently access or lower access frequency –miss penalty (server latency, bandwidth) Cache whole file or none Unicast Ignore limited disk bandwidth

9. Interval Caching [DaSi93, KaRT95] Resource Based Caching (RBC) [TVDS98] Least Frequently Used (LFU) Block-based insertion and deletion [AcSm00] Popularity-based caching for layered encoding [RYHE00] Prefix and Segment Caching for smoothing [SeRT99,WZDS98] Previous SM Caching Policies

10. Interval Caching Cache smallest intervals Target: memory caches (lots of insertions) File f 0 T Time   S1S1 S2S2 0 T S1S1 S1S1 S2S2 0 T  S1S1 S2S2 S3S3 0 T 

11. C ache entire files and intervals/runs Goal: efficiently utilize the limited resource –limited space: cache smallest space requirement –limited bandwidth: cache smallest write overhead Pre-allocate bandwidth to each cached entity Complex algorithm –Complex implementation –High time complexity Resource Based Caching

12. RBC Algorithm Step 1: Selecting entity x  {interval, run, file} of file i 1) If U bw > U space +  Choose the entity with lowest 2) If U space > U bw +  Choose the entity with minimum space requirement S i,x 3) If U space -  < U bw < U space +  Choose the entity with largest Step 2: Caching decision for entity x 1) If enough unallocated space and unallocated bandwidth: Cache entity x 2) If enough unallocated space but bandwidth constrained: Use bandwidth goodness list to select candidates for eviction 3) If enough unallocated bandwidth but space constrained: Use space goodness list to select candidates for eviction 4) If both bandwidth and space constrained: Walk on both lists: at each step, remove entity from bandwidth goodness list or from space goodness list. Step 3: Allocate space and bandwidth for entity x

13. Least Frequently Used Different implementation options: –What to do when receive first access to an object? –How to estimate frequency? Version studied: Currently Most Popular (CMP) –Insert only most frequently accessed (file or segment) –On-line popularity estimate: future research

14. Previous comparison : RBC vs. CMP [TVDS98] Fixed file access frequencies RBC outperforms CMP for all parameter values studied Limited design space –e.g.: total cache size  16GB Inconsistent results

15. New Performance Comparison Re-evaluate byte hit ratio of CMP and RBC –Simulation with synthetic workload –Broad design space New Pooled RBC New simple hybrid CMP/interval caching (CMP/IC) policy

16. System Assumptions Arrivals: Poisson( ) –extra experiments with Pareto( ,k) File access frequency: Zipf(  ) Perfect File popularity –extra experiments with approximate file popularity Uniform file size and delivery rate –extra experiments with variable file size and delivery rate Load balanced across multiple disks

17. System Parameters n : number of files  : Zipf parameter N : arrival rate (avg. number of requests per avg. file duration T) N =  T C : cache size (fraction of media data accessed)

18. B: normalized disk bandwidth (fraction of the average number of simultaneous streams needed to deliver data that is cached by CMP) B depends on N, , n, C and disk technology Relative performance of policies depends mainly on B B = 1.0 : CMP system is bandwidth balanced B  1.0 : CMP system is bandwidth deficient B  1.0: CMP system is bandwidth abundant System Parameters

19. Ultrastar 72ZX disk : –disk space: 116.76 hours of MPEG-1 video (73.4GB) –disk bandwidth: 108 MPEG-1 streams (22-37 MB/s ) Assume: 100 requests / hour for cached files If cache contains 2-hour movies: –Need 200 streams –B = 108/200 = 0.54 If cache contains 30-minute TV shows: –Need 50 streams for cache content –B = 108/50 = 2.16 Normalized Disk Bandwidth (B) Example

20. RBC vs. CMP CMP outperforms RBC if B  1.0 RBC slightly outperforms CMP if B  1.0 and small caches N = 450, n= 100,  =0

21. B = 0.75B = 2.0B = 1.0 Files Cached by RBC Average fraction of each file cached by RBC (N = 450, n = 100, C=0.25)

22. B = 0.75B = 2.0B = 1.0 Space and Bandwidth Utilization

23. Pooled RBC Three improvements over RBC –simpler rule to select entity to cache –can keep cached intervals when deleting a full file –pool of pre-allocated bandwidth Similar complexity as RBC

24. Pooled RBC, RBC and LFU Pooled RBC  CMP BUT, Pooled RBC is much more complex than CMP N = 450, n= 100,  =0

25. Hybrid CMP/IC Policies Do interval caching on a separate (small) cache –Interval Cache in Main Memory: CMP/IC mem and Pooled RBC/IC mem –Interval Cache on Disk: CMP/IC disk e.g. 5% of disk cache

26. N = 450, n= 100,  =0 CMP/IC mem vs. Pooled RBC/IC mem Memory cache improves CMP and Pooled RBC B  1.0 : greater improvement for CMP

27. N = 450, n= 100,  =0 CMP/IC disk vs. Pooled RBC CMP/IC disk  Pooled RBC  CMP

28. Conclusions Simple CMP –simple to implement –performance similar to Pooled RBC, CMP/IC disk (static file popularities) Hybrid CMP/IC policy –Performance  Pooled RBC –simple to implement –possibly more robust (imperfect and dynamic popularity measures)

29. Future Work Develop on-line estimate of file popularity Server log analysis –client behavior and workloads (NOSSDAV’01 paper) –More logs!!!! Caching Policies for Multicast Streams –popular file has greater cache-sharing if not cached –determine cache content that minimizes per-client cost –caching principles / on-line policy –(coming up soon) Prototype, experimental ( live ) workloads