Correcting the Dynamic Call Graph Using Control Flow Constraints

Transcription

1 Correting the Dynmi Cll Grph Using Control Flow Constrints Byeongheol Lee Kevin Resnik Mihel Bond Kthryn MKinley 1 Appered in CC2007

2 Motivtion Complexity of lrge objet oriented progrms Deompose the progrm into smll methods Method boundry beomes performne-bottlenek Dynmi interproedurl optimiztion Solve the method boundry problem Inlining nd speiliztion vry the performne by ftor of 2 Dynmi ll grph (DCG) is ritil input! b w 1 2 w 2 Dynmi ll grph

3 Inurte ll grph 1,000 b ll b ll 500 Error DCG Smple method 3

4 Timer-bsed smpling nd timing bis Cll stk b b b b t 4

5 Timer-bsed smpling nd timing bis Cll stk b b b b t 5

6 Timer-bsed smpling nd timing bis Cll stk b b b b t 6

7 Timer-bsed smpling nd timing bis Cll stk b b b b t 7

8 Timer-bsed smpling nd timing bis timer tik timer tik timer tik timer tik Cll stk b b b b t DCG Smple b b b b 8

9 Overhed nd ury in ll grph profiling 25 Full instrumenttion Overhed (%) Arnold-Grove smpling [2005] Timer-bsed smpling [2000] Aury (%) Corretion [2007] 100

10 Outline Motivtion Cll grph orretion Evlution 10

11 Timing bis in SPEC JVM98 rytre Smpling Normlized frequeny(%) Method lls grouped by soure method 11

12 Timing bis in SPEC JVM98 rytre Normlized frequeny(%) Method lls grouped by soure method 12

13 Corretion lgorithms Detet nd orret DCG error DCG onstrint Stti nd dynmi pprohes New Stti FDOM (Frequeny domintor) orretion Stti pproh Uses stti FDOM onstrint on DCG Dynmi bsi blok profile orretion Dynmi pproh Uses dynmi bsi blok profile onstrint on DCG 13

14 Stti FDOM onstrint FDOM onstrint on CFG ll is exeuted t lest s mny times s ll b ll FDOM ll b FDOM onstrint on DCG f( ) f( b ) ll b ll method 14

15 Stti FDOM orretion FDOM onstrint: f( ) f( b ) 1,000 b Corretion 750 b DCG Smple DCG FDOMCorretion 15 Detet error nd ssign the sme verge frequeny One possible solution to the FDOM onstrint Preserve totl frequeny sum

16 Dynmi bsi blok profile onstrint Some dynmi optimiztion systems do edge profiling Bseline ompiler in Jikes RVM 16 Dynmi bsi blok profile onstrint on CFG f(ll ) = 2 * f(ll b) Dynmi bsi blok profile onstrint on DCG f( ) = 2 * f( ) b method 50% 50% ll b ll

17 Dynmi bsi blok profile orretion Constrint: f( ) = 2* f( b ) 1,000 b Corretion 500 b 500 1,000 DCG Smple DCG EdgeProfileCorretion 17 f New ( b ) = 1/(1+2) * (1, ) = 500 f New ( ) = 2/(1+2) * (1, ) = 1,000

18 18 Best result: rytre 5 Normlized frequeny(%) Normlized frequeny(%) Normlized frequeny(%) Smpling Stti FDOM orretion 0 Dynmi bsi blok profile orretion

19 Outline Motivtion Cll grph orretion Evlution 19

20 Experimentl methodology Jikes RVM on 3.2G Pentium 4 Reply methodology [Blkburn et l. 06] Deterministi run 1 st itertion ompiltion + pplition run 2 nd itertion pplition run Mesurement Aury Use overlp ury [Arnold & Grove 05] Overhed 1 st itertion inludes ll grph orretion Performne 2 nd itertion is pplition-only SPECJVM98 nd DCpo benhmrks 20

21 21 Aury ompress jess Aury(%) rytre db jv mpegudio mtrt jk ntlr blot fop hsqldb jython luindex ipsixql jbb Averge No orretion Stti FDOM orretion Dynmi bsi blok profile orretion

22 22 Overhed ompress jess rytre Normlized exeution time db jv mpegudio mtrt jk ntlr blot fop hsqldb jython luindex ipsixql jbb Averge Stti FDOM Corretion Dynmi bsi blok profile orretion

23 23 Inlining performne Stti FDOM Corretion Dynmi bsi blok Profile orretion Perfet DCG ompress jess rytre db jv mpegudio mtrt jk ntlr blot fop hsqldb jython luindex ipsixql jbb Averge Normlized exeution time Bseline: profile-guided inlining with defult ll grph smpling

24 Summry CFG onstrint improves the DCG Inlining hs been tuned for bd ll grph Advntges Cn be esily ombined with other DCG profiling Miniml overhed only during the ompiltion Future work More inter-proedurl optimiztions with high ury DCG 24

25 Question nd omment Thnk you! 25

26 26

27 27

28 28

29 29

30 Timing bis misleds optimizer 5,000 times 10,000 times b Smpling with timing bis 1,000 smples 500 smples b DCG Perfet DCG Smple 30 DCG Smple Edge frequenies were reversed! Inlining deision Inliner my inline b insted of

31 Cll grph profiling in online optimiztion system Soure progrm e.g. Jv byte ode Compile & instrument Dynmi ll grph Mhine ode Online optimiztion system 31 Profiling nd progrm run t the sme time Minimize profiling overhed Corollry: srifie profiling ury