Driving Cache Replacement with ML-based LeCaR

Transcription

1 1 Driving Cache Replacement with ML-based LeCaR Giuseppe Vietri, Liana V. Rodriguez, Wendy A. Martinez, Steven Lyons, Jason Liu, Raju Rangaswami, Ming Zhao, Giri Narasimhan Florida International University Arizona State University

2 2 Reinforcement Learning On Cache Replacement Algorithms CACHE

3 Outline 3 Previous work Algorithm Results Conclusion

4 4 Hit-rate Performance vs. ARC Cache Size Non-Parameterized Adaptive Fixed-Parameter LRU LFU FBR LIRS MQ ARC LRU-2 2Q LRFU , Worse than ARC Better than ARC Megiddo, N., & Modha, D. S. (2004). Outperforming LRU with an adaptive replacement cache algorithm. Computer, 37(4),

5 Adaptive Replacement Cache (ARC) 5 p T 1 T 2 B 1 B 2 Megiddo, N., & Modha, D. S. (2004). Outperforming LRU with an adaptive replacement cache algorithm. Computer, 37(4),

6 6 Strengths of ARC 6 Manages both recent items as well as frequent items Dynamically adapts Self-tuning Low overhead Competitive Hit-Rate performance

7 Conventional Online Learning Systems 7 1. Choose Expert 2. Follow Advice W 1 3. Adjust Weights 4. Repeat Ai W 2 W 3 Ari, I., Amer, A., Gramacy, R.B., Miller, E.L., Brandt, S.A. and Long, D.D., 2002, March. ACME: Adaptive Caching Using Multiple Experts. In WDAS (pp ).

8 Not efficient 8 Disadvantages of current ML-based methods Not competitive

9 9 ML-based LeCaR LRU and LFU Efficiency Learning the optimal mix.

10 The LeCaR approach 10 LRU and LFU History Regret LeCaR Cache C LRU W LRU History H LRU H LFU LFU W LFU Lee, D., Choi, J., Kim, J. H., Noh, S. H., Min, S. L., Cho, Y., & Kim, C. S. (2001). LRFU: A spectrum of policies that subsumes the least recently used and least frequently used policies. IEEE transactions on Computers, (12),

11 Input: requested page q If q in C then C.Update(q) else if q is in HLRU then H LRU.Delete(q) if q is in H LFU then H LFU.Delete(q) UpdateWeights(q) if C is full then action = (LRU, LFU)~(w LRU,w LFU ) if act == LRU then if H LRU is full then H LRU.Delete(LRU(H LRU )) H LRU.Add(LRU(C)) C.Delete(LRU(C)) else if H LFU is full then H LFU.Delete(LFU(H LFU )) H LFU.Add(LFU(C)) C.Delete(LRU(C)) C.Add(q) LeCaR (Hit) Request Weights Cache History W_lru W_lfu H lru H lfu

12 Input: requested page q If q in C then C.Update(q) else if q is in HLRU then H LRU.Delete(q) if q is in H LFU then H LFU.Delete(q) UpdateWeights(q) if C is full then action = ( LRU, LFU)~(w LRU,w LFU ) if act == LRU then if H LRU is full then H LRU.Delete(LRU(H LRU )) H LRU.Add(LRU(C)) C.Delete(LRU(C)) else if H LFU is full then H LFU.Delete(LFU(H LFU )) H LFU.Add(LFU(C)) C.Delete(LRU(C)) C.Add(q) LeCaR (Miss not in History) Request Cache History H lru H lfu 12 W_lru W_lfu 1 8

13 Weight Update 13 α = Learning rate LRU Regret LFU Regret Update w LRU := α * w LRU w LFU := α * w LFU Normalize w LRU := w LRU /(w LRU + w LFU ) w LFU := 1 w LRU

14 Input: requested page q If q in C then C.Update(q) else if q is in HLRU then H LRU.Delete(q) if q is in H LFU then H LFU.Delete(q) UpdateWeights(q) if C is full then action = ( LRU, LFU)~(w LRU,w LFU ) if act == LRU then if H LRU is full then H LRU.Delete(LRU(H LRU )) H LRU.Add(LRU(C)) C.Delete(LRU(C)) else if H LFU is full then H LFU.Delete(LFU(H LFU )) H LFU.Add(LFU(C)) C.Delete(LRU(C)) C.Add(q) LeCaR (Miss in History) Request Weights Cache History H lru W_lru W_lfu 7 20 H lfu

15 Experiments 15 8 Workloads 3 days of data[fiu datasets] Small, Medium and Large Cache Sizes Fixed Learning Rate

16 Data used for the Experiments 16 Casa, Ikki, Madmax and Topgun Four different end-user/developer home directories Online Departments online course management system Webresearch Document store for research projects Webmail Webuser Mail server of the FIU Computer and Engineering department using Postfix Web server hosting faculty, staff, and graduate student web sites Collected at the School of Computing and Information Sciences at FIU. R. Koller, R. Rangaswami. I/O deduplication: Utilizing content similarity to improve i/o performance. In Proc. 8th USENIX Conference on File and Storage Technologies (2009), FAST

17 Online 75 Results % 0.50% 1% 5% 10% LRU LFU ARC LeCaR

18 Results LRU ARC LFU LeCaR 18

19 Results LRU ARC LFU LeCaR 19

20 Synthetic Data 20

21 Synthetic Data 21

22 Hoarding Rate Hoarded Page: Accessed at least twice Not among the last 2N unique pages 22 Definition A: Stable Period B: LFU gets penalized

23 Conclusion 23 ML-based LeCaR Small cache size Real + Synthetic experiments