Papers We Love: ARC after dark

My Papers We Love talk in San Francisco on October 12, 2017 on "ARC: A self-tuning, low overhead replacement cache." Video: https://www.youtube.com/watch?v=F8sZRBdmqc0

Transcript

1. Papers We Love:
   ARC after dark
   CTO
   bryan@joyent.com
   Bryan Cantrill
   @bcantrill
   

   View Slide

2. Papers we love: ARC
   • Discussing ARC: A self-tuning, low overhead replacement
   cache by Nimrod Megiddo and Dharmendra Modha
   • Appeared in USENIX FAST ’03
   • Has been broadly deployed in production — most notably in
   ZFS, in at least illumos (SmartOS, etc.) and FreeBSD
   • As much as possible, want to let the paper speak for itself —
   and provoke discussion!
   

   View Slide

3. ARC in the abstract
   

   View Slide

4. Introduction
   

   View Slide

5. Book of Genesis, 1:1
   

   View Slide

6. The Von Neumann architecture
   

   View Slide

7. “The Memory Organ”
   

   View Slide

8. “Entirely impractical”
   

   View Slide

9. The problem
   

   View Slide

10. Main themes
    

    View Slide

11. Related work: Optimal
    

    View Slide

12. Related work: Using recency with LRU
    

    View Slide

13. Related work: Using frequency with LFU
    

    View Slide

14. Related work: LRU-2
    

    View Slide

15. Related work: LRFU
    

    View Slide

16. Related work: ALRFU (AYFKM?!)
    

    View Slide

17. Related work: MQ
    

    View Slide

18. Related work: MQ, cont.
    

    View Slide

19. Related work: “Multiple experts”
    

    View Slide

20. Related work: Ghost caches
    

    View Slide

21. Generic cache replacement policy (?)
    

    View Slide

22. Fixed replacement cache
    

    View Slide

23. Adaptive replacement cache
    

    View Slide

24. Adaptive replacement cache
    

    View Slide

25. Adaptive replacement cache
    

    View Slide

26. Adaptive replacement cache
    

    View Slide

27. Adaptive replacement cache
    

    View Slide

28. Adaptive replacement cache
    

    View Slide

29. Adaptive replacement cache
    

    View Slide

30. ARC performance
    

    View Slide

31. ARC performance
    

    View Slide

32. ARC performance
    

    View Slide

33. ARC performance vs. LRU
    

    View Slide

34. ARC performance vs. FRC offline
    

    View Slide

35. ARC performance: “Empirically universal”
    

    View Slide

36. ARC performance: p over time
    

    View Slide

37. ARC: Conclusions
    

    View Slide

38. ARC: Endnotes
    

    View Slide

39. ARC: Endnotes
    

    View Slide

40. ARC: Epilogue
    • The ARC is the basis of the cache used by ZFS, with many
    enhancements over time:
    • Variable sized buffers — though now returning to be fixed
    blocks with ARC buffer data (ABD)
    • L2ARC to allow SSDs to be used as a cache
    • Separate metadata lists
    • Compressed ARC
    • A challenge has been to make all of ARC variable-sized…
    

    View Slide

41. Peril of a variably-sized ARC!
    

    View Slide