1. 1 Storage-Aware Caching: Revisiting Caching for Heterogeneous Storage Systems (Fast’02) Brian Forney Andrea Arpaci-Dusseau Remzi Arpaci-Dusseau Wisconsin Network Disks University of Wisconsin at Madison Presented by Enhua Tan July 1, 2005

2. 2 Circa 1970 OS policy Intense research period in policies Wide variety developed; many used today –Examples: Clock, LRU Simple storage environment –Focus: workload –Assumption: consistent retrieval cost buffer cache App

3. 3 Today Rich storage environment –Devices attached in many ways –More devices –Increased device sophistication Mismatch: Need to reevaluate LAN WAN policy buffer cache App

4. 4 Problem illustration Uniform workload Two disks LRU policy Slow disk is bottleneck Problem: Policy is oblivious –Does not filter well fastslow

5. 5 General solution Integrate workload and device performance –Balance work across devices –Work: cumulative delay Cannot throw out: –Existing non-cost aware policy research –Existing caching software

6. 6 Existing cost-aware algorithm LANDLORD –Cost-aware, based on LRU / FIFO –Same as web caching algorithm –Integration problems with modern OSes

7. 7 LANDLORD (LRU version) Each object associates with a cost L –Initially set to H: the retrieval cost divided by the size of the object –Evict the object with lowest L; and decrementing all remaining pages by this L –Upon reference, L restored to H When all H values are the same, equal to strict LRU

8. 8 Solution Overview Generic partitioning framework –Old idea –One-to-one mapping: device  partition –Each partition has cost-oblivious policy –Adjust partition sizes Advantages –Aggregates performance information –Easily and quickly adapts to workload and device performance changes –Integrates well with existing software Key: How to pick the partition sizes

9. 9 Outline Motivation Solution overview Taxonomy Dynamic partitioning algorithm Evaluation Summary

10. 10 Partitioning algorithms Static –Pro: Simple –Con: Wasteful – cannot balance work for all workloads Dynamic –Adapts to workload Hotspots Access pattern changes –Handles device performance faults Used dynamic

11. 11 Algorithm: Overview 1.Observe: Determine per-device cumulative delay 2.Act: Repartition cache 3.Save & reset: Clear last W requests Observe Act Save & reset

12. 12 Algorithm: Observe Want accurate system balance view Record per-device cumulative delay for last W (window size) completed disk requests –At client –Includes network time (and the remote disk service time)

13. 13 Algorithm: Act Categorize each partition –Page consumers Cumulative delay above threshold  possible bottleneck –Page suppliers Cumulative delay below mean  lose pages without decreasing performance –Neither Always have page suppliers if there are page consumers Page consumerPage supplierNeither Page consumerPage supplierNeither Before After

14. 14 Page consumers How many pages? Depends on state: –Warming Cumulative delay increasing Aggressively add pages: reduce queuing –Cooling Cumulative delay decreasing Do nothing; naturally decreases –Warm Cumulative delay constant Conservatively add pages

15. 15 Three parameters W: window size –W = 1000 provides sufficient smoothing and feedback T: threshold –T = 5, can detect changes in relative wait time that need correction I: base increment amount –Should be small for exponential correction –0.2% of the cache works well

16. 16 Dynamic partitioning Eager –Immediately change partition sizes –Pro: Matches observation –Con: Some pages temporarily unused Lazy –Change partition sizes on demand –Pro: Easier to implement –Con: May cause over correction

17. 17 Partitioned-Clock algorithm Each page has a use bit like the Clock algorithm Also has the partition number that it belongs Victim will be selected with the matched partition number for replacement A separate clock hand for each partition helps

18. 18 Lazy Clock When searching for a replacement –Only those pages belonging to the victim’s partition should have their use bits cleared Uses inverse lottery scheduling to pick victims among partitions –Each partition is given a number of tickets in inverse proportion to its desired size

19. 19 Outline Motivation Solution overview Taxonomy Dynamic partitioning algorithm Evaluation Summary

20. 20 Evaluation methodology Simulator Workloads: synthetic and web Caching, RAID-0 (non-redundant) client LogGP network With endpoint contention Disks 16 IBM 9LZX (9.1GB; SCSI; 10,020 RPM) First-order model: queuing, seek time & rotational time

21. 21 LogGP Model 0.5μs 0.3μs 76ns 1/G=10 Gb/s

22. 22 Evaluation methodology Introduced performance heterogeneity –Disk aging Used current technology trends –Seek and rotation: 10% decrease/year –Bandwidth: 40% increase/year Scenarios –Single disk degradation: Single disk, multiple ages –Incremental upgrades: Multiple disks, two ages –Fault injection Understand dynamic device performance change and device sharing effects

23. 23 Evaluated policies Cost-oblivious: LRU, Clock Storage-aware: Eager LRU, Lazy LRU; Lazy Clock, Eager Clock Comparison: LANDLORD

24. 24 A motivating example LRU: 55MB/s -> 11MB/s

25. 25 Static partitioning Statically set: the ratio of partition sizes ∞ the ratio of the expected service time of each disk

26. 26 Synthetic Workload: Read requests, exponentially distributed around 34 KB, uniform load across disks A single slow disk greatly impacts performance. Eager LRU, Lazy LRU, Lazy Clock, and LANDLORD robust as slow disk performance degrades

27. 27 Multiple old disks

28. 28 Dynamic fault Relatively stable partition sizes found within 75 seconds Eager LRU/Clock have the highest bandwidth

29. 29 Web Workload: 1 day image server trace at UC Berkeley, reads & writes Eager LRU and LANDLORD are the most robust.

30. 30 Web without writes Workload: Web workload where writes are replaced with reads Eager LRU and LANDLORD are the most robust.

31. 31 Summary Problem: Mismatch between storage environment and cache policy –Current buffer cache policies lack device information Policies need to include storage environment information Solution: Generic partitioning framework –Aggregates performance information –Adapts quickly –Allows for use of existing policies

32. 32 Limitations Assumptions –Linear relationship between cache size and hit rate Doesn’t hold for workload with little locality –Rely on proper values for three parameters: window size, base increment amount, and threshold Outperform dramatically when a disk is quite slow (4~10 years old) –But is that a realistic problem with technology advancing? Should the disk be replaced rather than using more cache space to mask the lag? Break transparency: –Clients need to know there exist 16 network disks?

33. 33 Future work Implementation in Linux Cost-benefit algorithm Study of integration with prefetching and layout

34. 34 Follow-ups 10 citations found in Google Scholar, most are before 2003 Papers in the field of Web, Cluster, and Storage system

35. 35 WiND project Wisconsin Network Disks Building manageable distributed storage Focus: Local-area networked storage Issues similar in wide-area