Download Whole Execution Trace and Its Applications in CS 260.2 and more Study notes Computer Science in PDF only on Docsity!
Whole Execution Trace and its
Applications
CS 260.
2 Execution Histories Control Flow Dependences Exercised Addresses Referenced Values Computed Compiler Optimizations Hot Path Specialization Value-based Specialization Locality Improving Transformations Data Speculation etc. Profile-Guided Redundant and Dead Code Elimination Branches [ PLDI’97a ] Assignments [ PLDI’97b ] Expressions [ MICRO’97,PLDI’ 98 ] Loads/Stores [ PLDI’99,IJPP’ 96 ] Bounds Checks [ PLDI’ 00 ] 3 Execution Histories Control Flow Dependences Exercised Addresses Referenced Values Computed Processor Architectures Branch Prediction Value Prediction Code & Data Prefetching Load-Store Disambiguation Power Frequent Values [ ASPLOS’00,ACM TECS’ 02 ] Compression Caches [ MICRO’02&’ 00 ] Off-Chip Buses [ ACM TODAES’ 04 ] Performance Value Prediction [ HPCA’99,ISCA’ 99 ] Load-Store Disambiguation [ MICRO’ 99 ] 4 Execution Histories Control Flow Dependences Exercised Addresses Referenced Values Computed Debugging & Testing Delta Debugging Dynamic Slicing Likely Invariants Test Suite Coverage Dynamic Matching Comparing executions of two versions [ ESEC/FSE -05 ] Efficient - space and time [ ICSE-03, TOPLAS 2005, ICSE-04, PLDI-04 ] Effective in locating faults [ AADEBUG-05, ASE-05, ICSE-06, PLDI-06 ] Dynamic Slicing 5 Whole Execution Trace Representation Static Program Representation
- Control flow graph
- Program Dependences Data and control dependences Dynamic Profile Representation
- Designed for Analysis Annotates static program representation Related information can be easily accessed
- Comprehensive Control flow Addresses and values Data and control dependences
- Compact Design compression techniques 6 Input: N= Representation of Dynamic Information 51 : for I=1 to N do 61 : if (i%2==0) then 71 : p=&a 81 : a=a+ 91 : z=2(p) 101 : print(z) 11 : z= 21 : a= 31 : b= 41 : p=&b 52 : for I=1 to N do 62 : if (i%2==0) then 82 : a=a+ 92 : z=2(p) 1: z= 2: a= 3: b= 4: p=&b 5: for i = 1 to N do 6: if ( i %2 == 0) then 7: p=&a endif endfor 8: a=a+ 9: z=2(p) 10: print(z) &b &a 0 0 2 1 F 1 4 2 T 2 4 4
7 5:for i=1 to N 6:if (i%2==0) then 7: p=&a 8: a=a+ 9: z=2(p) 10: print(z) T** (^) F 1: z= 2: a= 3: b= 4: p=&b T Input: N= 11 : z= 21 : a= 31 : b= 41 : p=&b 51 : for i = 1 to N do 61 : if ( i %2 == 0) then 81 : a=a+ 91 : z=2(p) 52 : for i = 1 to N do 62 : if ( i %2 == 0) then 71 : p=&a 82 : a=a+ 92 : z=2(p) 101 : print(z)
T 1 2 3 4 5 6 7 8 9
10 11 12 13 14 Annotating Static Rep. With Dynamic Info. (2,7) (3,8) (7,12) (11,13) (13,14) (4,8) (12,13) (5,6)(9,10) (10,11) (5,7)(9,12) (5,8)(9,13) 1 2 3 4 5, 6, 11 7, 8, 14 F &b &a 0 0 2 1, F,T 1, 4, 4 8 WET Representation < tn > Block n^ Basic S 1 S 2 ... < v 1 > < v 2 > … < (ts, tn) > <^ (ts,^ tn) > 9 WET Sizes 10, 11, 8, 8, 10, 9, 10, 5, 10, 9,
300.twolf 256.bzip 255.vortex 197.parser 181.mcf 164.gzip 130.li 126.gcc 099.go Average
WET
Representation Size (MB) Statements Executed (Millions) Program 10 Compaction of WET Representation
Custom Compaction
Node o Timestamps o Values/Addresses Dependence Edges o Timestamp pairs
Generic Compression
All sequences Context-based method 11 Node Timestamps 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 2
3 (^46) 7 9 3 (^56) 7 9 3 (^46) 8 9 3 (^56) 8 9 1 2 3
3 (^46) 7 9 1 2 3 3 4 5 6 7 8 9 12 Infer: Local Dependence Labels: Full Elimination X = Y= X
X =
Y= X
(10,10) (20,20) (30,30)
X =
Y= X
10,20, =Y 21 (20,21) ... (...,20) ...
19 Node Values Pattern = 010121 ababcb dedefe mnmnon pqpqrq ghghih jkjklk abc (^) def mno pqr ghi jkl 20 Generic Compression 1 X Y Z 1 X Y Z A X Y Z B AA X Y ZX Y Z B
Traversable
Bidirectional
- Forward traversal
- Backward traversal 21 Compacted WET Sizes 690 751 609 615 715 650 740 365 685 647 Statements Executed (Millions) Before After
300.twolf 256.bzip 255.vortex 197.parser 181.mcf 164.gzip 130.li 126.gcc 099.go Average Before / After WET Size (MB) Program 22 Customized vs. Generic Customized x Generic = Total = 16. = 54. = 83. = 46. = 25. = 18. = 51. = 58. = 18. = 41. x 2. x 5. x 6. x 6. x 4. x 3. x 5. x 5. x 2. x 4.
300.twolf 256.bzip 255.vortex 197.parser 181.mcf 164.gzip 130.li 126.gcc 099.go Average Program 23 Scalability 0 10 2030 40 5060 70 80 Execution length Compaction factor 099.go 126.gcc 130.li 164.gzip 181.mcf 197.parser 255.vortex 256.bzip2 300.twolf 1 Gigabyte ⇔ 2 Billion Statements **0 1020 3040 50 60 7080 90 Execution length Compaction factor 099.go 126.gcc 130.li 164.gzip 181.mcf 197.parser 255.vortex 256.bzip2 300.twolf (billion)
24** Dependences Eliminated
25 Architecture Specific Information
300.twolf 256.bzip 255.vortex 197.parser 181.mcf 164.gzip 130.li 126.gcc 099.go Cache Hit/Miss History (MB) Branch Outcomes (MB) Program 26 Response Time: Control Flow Query
300.twolf 256.bzip 255.vortex 197.parser 181.mcf 164.gzip 130.li 126.gcc 099.go Program Time (seconds) 27 Response Time: Load Value/Address Query
300.twolf 256.bzip 255.vortex 197.parser 181.mcf 164.gzip 130.li 126.gcc 099.go Program Load Values
Addresses 28 Dynamic Slicing
Dynamic slice is the set of statements that DID
affect the value of a variable at a program point
for ONE specific execution. [Korel and Laski, 1988]
- Execution history dynamic control dependences dynamic data dependences
- Construct a dynamic dependence graph
- Traverse dynamic dependence graph to compute slices
- Smaller, more precise, slices are more helpful 29 Dynamic Data Slice ……
- A = …...
- B = ……
- P =
- If (P<0) {
- A = A + 1
- B = B + 1
- } ……
- Error(A+B) 30 Dynamic Full Slice ……
- A = …...
- B = ……
- P =
- If (P<0) {
- A = A + 1
- B = B + 1
- } ……
- Error(A+B) ……
- A = …...
- B = ……
- P =
- If (P<0) {
- A = A + 1
- B = B + 1
- } ……
- Correct(A+B)
37 Matching Program Versions
Matching produces a MAPPING between executed
instructions belonging to two versions of a program
by comparing their WETs for the same input.
Debugging Optimized Code
Cause of erroneous behavior?
Bug in the optimizer; or Bug exposed by the optimizer. 38 Comparison Checking [ESEC/FSE 1999]
Drawbacks
- Compiler writer must produce mappings
- Compiler modification may not be possible Source Code mapping Execute Execute
Compiler
OPT
UNOPT
Compare
trace trace Report 39 Matching Overview
Instruction Signatures
- Exact Match versus Consistency
Matching Dynamic Data Dependence Graph
- Matching Roots Order according to first execution
- Matching Other Nodes Use graph structure
Matching Possibilities
- No matches, One-to-one matches, & Many-to-one matches
Matching Accuracy
- Missed matches
- False matches 40 Unoptimized vs. Optimized Versions
li.U li.O m88ksim.U m88ksim.O twolf.U twolf.O go.U go.O parser.U parser.O Dynamic Program Static (millions) 41 Matches Generated 302 289 362 387 265 Time (sec) 97.0 % 95.1 % 94.2 % 91.0 % 96.1 % li m88ksim twolf go parser Program Optimized 42 Matching Accuracy 0 4 0 0 0 Missed 35 72 123 27 29 False 112 823 1067 1362 243 li m88ksim twolf go parser Program Actual
43 Comparison Checking 44 Applications of Version Matching
Debugging Optimized Code
- Comparing unoptimized and optimized versions
Software Piracy Detection
- Measuring the degree of similarity in the computation
Others…
45 Whole Execution Trace Representation
Static and Dynamic Information
- Definite & likely events can be queried
Design for Analysis
- Efficient analysis for answering queries
Compact Representation
- Sufficient dynamic information can be kept
Comprehensive
- Wide range of queries can be answered
Extensible
- New types of information/queries can be added 46 Work In Progress and Future Work
Debugging, Slicing and Matching
- Dynamic Slicing for Multiple runs/inputs
- Dynamically altering branch outcomes
- Code Obfuscation
Making Execution Histories Available
- API – Compaction is transparent
- Support for range of queries
Other forms of Dynamic Histories
- Architecture specific information
- Storage allocation/deallocation
- System calls
- Multiple Runs
- Multithreading
