CLIC CLient-Informed Caching for Storage Servers

Transcription

1 CLIC CLient-Informed Caching for Storage Servers Xin Liu Ashraf Aboulnaga Ken Salem Xuhui Li David R. Cheriton School of Comuter Science University of Waterloo February 2009

2 Two-Tier Caching DBMS storage server

3 Two-Tier Caching DBMS 1. read() storage server

4 Two-Tier Caching DBMS 1. read() storage server 2. read()

5 Two-Tier Caching DBMS 1. read() storage server 2. read() 3. fetch

6 Two-Tier Caching DBMS 1. read() 4. fetch storage server 2. read() 3. fetch

7 Two-Tier Caching DBMS 1. read() 4. fetch Problems: inclusion storage server 2. read() 3. fetch

8 Two-Tier Caching DBMS 1. read() 4. fetch Problems: inclusion storage server oor temoral locality 2. read() 3. fetch

9 Two-Tier Caching DBMS 1. read() 4. fetch Problems: inclusion storage server oor temoral locality One Solution: hinting 2. read() 3. fetch

10 Examle: Write Hints DBMS write() storage server

11 Examle: Write Hints DBMS write() this is a relacement write storage server

12 Examle: Write Hints DBMS write() storage server this is a relacement write this is a good candidate for caching

13 Examle: Write Hints DBMS write() storage server this is a relacement write this is a good candidate for caching The storage server can use TQ, an ad hoc hint-aware relacement olicy, to exloit write hints.

14 Problems with Ad Hoc Hint-Aware Policies narrowness: new hints? multile hints?

15 Problems with Ad Hoc Hint-Aware Policies narrowness: new hints? multile hints? brittleness: correct resonse to hints?

16 Problems with Ad Hoc Hint-Aware Policies narrowness: new hints? multile hints? brittleness: correct resonse to hints? single source: multile hint generators? DBMS DBMS q this is a relacement write write() storage server q write(q) this is a relacement write q should I or q????

17 The CLIC Aroach a hint-aware caching olicy for 2nd-tier s

18 The CLIC Aroach a hint-aware caching olicy for 2nd-tier s no hard coded resonse to secific hints

19 The CLIC Aroach a hint-aware caching olicy for 2nd-tier s no hard coded resonse to secific hints instead, learn which hints signal good caching oortunities

20 The CLIC Aroach a hint-aware caching olicy for 2nd-tier s no hard coded resonse to secific hints instead, learn which hints signal good caching oortunities benefits: handles multile hint tyes handles new hint tyes handles hints from multile clients by treating each client s hints as distinct

21 The CLIC Aroach a hint-aware caching olicy for 2nd-tier s no hard coded resonse to secific hints instead, learn which hints signal good caching oortunities benefits: handles multile hint tyes handles new hint tyes handles hints from multile clients by treating each client s hints as distinct CLIC Hints CLIC searates the generation of hints (done by the storage clients) from the interretation of those hints for caching uroses (done by the storage server).

22 CLIC Illustrated DBMS this is a blargh gor read read() storage server I don t know blargh or gor but revious blargh gor reads have been good candidates, so I will

23 Generating Hints Storage client must be modified to generate one or more tyes of hints. Storage clients attach a hint set to each read or write request. A hint set includes one hint of each tye generated by the client. A storage client may choose to generate any tyes of hints that might be of use to the storage server.

24 Generating Hints Storage client must be modified to generate one or more tyes of hints. Storage clients attach a hint set to each read or write request. A hint set includes one hint of each tye generated by the client. A storage client may choose to generate any tyes of hints that might be of use to the storage server. Examle: Hints from DB2 buffer ool ID object ID: identifies a grou of related DB objects object tye ID: distinguishes table from index request tye: read, relacement/recovery write DB2 buffer riority

25 A CLIC-Managed Cache highest riority Ha Hb each age is associated with the hint set which it was most-recently read or written Hc Hd He lowest riority

26 A CLIC-Managed Cache highest riority Ha Hb Hc each age is associated with the hint set which it was most-recently read or written each hint set has a riority Hd 6 He 8 lowest riority

27 A CLIC-Managed Cache highest riority Ha Hb Hc Hd each age is associated with the hint set which it was most-recently read or written each hint set has a riority CLIC evicts ages associated with the lowest-riority hint sets He 8 lowest riority

28 A CLIC-Managed Cache highest riority Ha 1 3 Hb 7 5 Hc 2 9 Hd 6 He 8 lowest riority 4 each age is associated with the hint set which it was most-recently read or written each hint set has a riority CLIC evicts ages associated with the lowest-riority hint sets CLIC chooses hint set riorities using a simle cost/benefit analysis

29 Cost/Benefit Analysis here?? time (,H) read or write request next request for

30 Cost/Benefit Analysis here?? time (,H) read or write request next request for is this a read request? there is a benefit to caching if the next request for is a read request

31 Cost/Benefit Analysis here?? time (,H) read or write request next request for is this a read request? there is a benefit to caching if the next request for is a read request the cost of obtaining this benefit is that must remain d until the read request

32 Assigning Priorities to Hint Sets here?? time (,H) read or write request next request for is this a read request? when request (, H) occurs, CLIC cannot know the the cost and benefit of caching

33 Assigning Priorities to Hint Sets here?? time (,H) read or write request next request for is this a read request? when request (, H) occurs, CLIC cannot know the the cost and benefit of caching instead CLIC estimates the cost and benefit of caching at (, H) based on revious requests with hint set H

34 Assigning Priorities to Hint Sets here?? time (,H) read or write request next request for is this a read request? when request (, H) occurs, CLIC cannot know the the cost and benefit of caching instead CLIC estimates the cost and benefit of caching at (, H) based on revious requests with hint set H CLIC assigns a riority to each hint set based on the cost and benefit of revious requests with hint set H Priority(H) = Read Hit Rate(H) Mean Time Until Read Hit(H)

35 DB2 Hint Analysis Examle STOCK table relacement writes ORDERLINE table reads

36 Efficient Hint Analysis To analyze the cost and benefit of hint sets, CLIC must track the most recent request and hint set for each age track the mean read hit rate and read hit distance for each hint set

37 Efficient Hint Analysis To analyze the cost and benefit of hint sets, CLIC must track the most recent request and hint set for each age track the mean read hit rate and read hit distance for each hint set To reduce sace requirements, CLIC tracks the most recent request only for d ages and a fixed number of additional, und aged

38 Efficient Hint Analysis To analyze the cost and benefit of hint sets, CLIC must track the most recent request and hint set for each age track the mean read hit rate and read hit distance for each hint set To reduce sace requirements, CLIC tracks the most recent request only for d ages and a fixed number of additional, und aged tracks read hit statistics only for frequently occurring hint sets

39 Efficient Hint Analysis To analyze the cost and benefit of hint sets, CLIC must track the most recent request and hint set for each age track the mean read hit rate and read hit distance for each hint set To reduce sace requirements, CLIC tracks the most recent request only for d ages and a fixed number of additional, und aged tracks read hit statistics only for frequently occurring hint sets We have also investigated the use of generalization to reduce the number of distinct hint sets.

40 Performance we have used trace-driven simulation of the storage server buffer to comare CLIC to other relacement olicies methodology 1. modify DB2 and MySQL to generate hints and roduce I/O traces 2. run TPC-C (on-line transaction rocessing) and TPC-H (decision suort) workloads on the database systems and collect I/O traces 3. feed the traces to a simulation of second-tier, which imlements CLIC, LRU, ARC, TQ and OPT 4. measure the hit ratio achieved by different olicies.

41 DB2 TPC-C - Medium DB2 Buffer Cache Server Cache Read Hit Ratio 100% 80% 60% 40% 20% OPT TQ LRU ARC CLIC DB2_C300 0% 60k 120k 180k 240k 300k Server Cache Size (ages)

42 DB2 TPC-H - Medium DB2 Buffer Cache Server Cache Read Hit Ratio 100% 80% 60% 40% 20% OPT TQ LRU ARC CLIC DB2_H400 0% 60k 120k 180k 240k 300k Server Cache Size (ages)

43 DB2 TPC-C - Small DB2 Buffer Cache 100% DB2_C60 Server Cache Read Hit Ratio 80% 60% 40% 20% OPT TQ LRU ARC CLIC 0% 60k 120k 180k 240k 300k Server Cache Size (ages)

44 DB2 TPC-C - Large DB2 Buffer Cache Server Cache Read Hit Ratio 100% 80% 60% 40% 20% OPT TQ LRU ARC CLIC DB2_C540 0% 60k 120k 180k 240k 300k Server Cache Size (ages)

45 Summary and Conclusions CLIC learns to identify I/O hints that signal good caching oortunities by tracking the request stream observed by the second-tier Because CLIC s resonses to secific hints are not redefined, it naturally accommodates new hint tyes and hints from multile storage clients. for our traces: CLIC s erformance usually dominates ARC s and LRU s, sometimes by a factor of 2 or more. CLIC dominates the ad hoc, hint-aware TQ algorithm CLIC s sace overhead can be ket low ( 1% of storage server size in our exeriments)