SRM-Buffer: An OS Buffer Management Technique to Prevent Last Level Cache from Thrashing in Multicores

Transcription

1 SRM-Buffer: An OS Buffer Management Technique to Prevent Last Level Cache from Thrashing in Multicores Xiaoning Ding et al. EuroSys 09 Presented by Kaige Yan 1

2 Introduction Background SRM buffer design Evaluation Related Work Conclusion & Discussion 2

3 Introduction OS buffer keeps recently accessed file blocks in memory and it causes performance degradation Weaker temporal locality LLC is shared in multicore Keep the pollution due to accessing OS buffer to a certain region in cache Detection of accessing to OS buffer Coordinate the buffer and virtual memory management 3

4 Introduction Background SRM buffer design Evaluation Related Work Conclusion & Discussion 4

5 Background Run grep and mergesort on dual core system Big performance degradation due to the a lot of OS buffer access of grep 5

6 Background (cont.) Page coloring can do the job Limit the pollution to some colors for buffer data 6

7 Background (cont.) Allocating pages with same color for buffer data cause uneven page color distribution Decrease the hit ratio of active pages which are reclaimed prematurely 7

8 Introduction Background SRM buffer design Evaluation Related Work Conclusion & Discussion 8

9 Detection of Accessing to OS Buffer File blocks are usually accessed sequentially The maximum sequence length is Tl Same-file Heuristic Same-application Heuristic 9

10 SRM Buffer Design Active pages and most of inactive pages are managed in Normal Zone Colored Zone manage free list and some inactive pages, the free list are On page fault or OS buffer miss, SRM buffer reclaim pages in the colored zone 10

11 Introduction Background Hot Page Coloring Evaluation Related Work Conclusion & Discussion 11

12 Experiment Setup Linux PowerEdge 1900 : Two Quad-core Intel Xeon GHz CPU + 16G memory Precision T1500: Intel Core i7 860 quad-core + 8G memory Tl =

13 Database Workload 1 (simple query) Fact table: 4G, dimension table M Task 1: hash-join of fact table and dimension table Task 2: sequential scan of fact table 13

14 Database Workload 2(complex query) TPC-H queries First group: Q6 and Q15(more sequential scan) second group: Q5,Q7,Q8,Q10,Q11 14

15 Other workloads Grep, tar, PostMark (emulate internet applications), mergesort, FFT, MM and LU VM intensive: mergesort, FFT, MM, LU File intensive: grep, tar and PostMark 15

16 Performance on PowerEdge 1900 The slowdown is reduced by 28% The cache misses are reduced 34%-85% 16

17 Experiment on Access Pattern Changes Warm up the OS buffer with the file-intensive app inside the bracket and co-run the file intensive app outside the bracket and the VM-intensive one Access pattern changes reduce the effectiveness 17

18 Experiment with Contrived Adverse Workload When the OS almost runs out of free pages, SRM buffering could be detrimental to the performance The hit ratio is reduced by 11% with Tl=256 and 3% with with Tl=128 18

19 Parameter Sensitivity Long Tl reduce the cache pollution but decrease the hit ratio Short Tl increase the cache pollution but increase the hit ratio 19

20 Introduction Background Hot Page Coloring Evaluation Related Work Conclusion & Discussion 20

21 Related Work The shared last level cache in multicore is important for the performance Scheduling policy: classify the VM-intensive and file-intensive is hard and it depends on the availability of CPU cores. [Zhuravlev 10] Cache partitioning: not work for pollution due to OS buffer and OS buffer is big. [Lu 09] 21

22 Introduction Background Hot Page Coloring Evaluation Related Work Conclusion & Discussion 22

23 Conclusion & Discussion This paper propose SRM buffer strategy to alleviate the cache pollution due to OS buffer access. Can this method be combined with page color based cache partitioning policies? 23