Slicing Concurrent Programs Achievements and Open Challenges Dennis - - PowerPoint PPT Presentation

▶
slicing concurrent programs
SMART_READER_LITE
LIVE PREVIEW

Slicing Concurrent Programs Achievements and Open Challenges Dennis - - PowerPoint PPT Presentation

Jan 30, 2023 487 likes •895 views

Slicing Concurrent Programs Achievements and Open Challenges Dennis Giffhorn Programming paradigms group IPD Snelting Interference dependence Concurrency via threads and shared memory Shared-memory communication gives rise to interference

slide-1
SLIDE 1

Programming paradigms group – IPD Snelting Dennis Giffhorn

Achievements and Open Challenges

Slicing Concurrent Programs

slide-2
SLIDE 2

Interference dependence

Concurrency via threads and shared memory Shared-memory communication gives rise to interference dependence

Definition

Statement t is interference-dependent on statement s, if t uses a value which is defined by statement s, and s and t may happen in parallel

2

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-3
SLIDE 3

Interference dependence

Concurrency via threads and shared memory Shared-memory communication gives rise to interference dependence

Definition

Statement t is interference-dependent on statement s, if t uses a value which is defined by statement s, and s and t may happen in parallel

Example

1 thread_1: 2 a = input + c; 3 b = input - 5; 4 thread_2: 5 c = b;

2

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-4
SLIDE 4

Interference dependence

Concurrency via threads and shared memory Shared-memory communication gives rise to interference dependence

Definition

Statement t is interference-dependent on statement s, if t uses a value which is defined by statement s, and s and t may happen in parallel

Example

1 thread_1: 2 a = input + c; 3 b = input - 5; 4 thread_2: 5 c = b;

2

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-5
SLIDE 5

Interference dependence

Concurrency via threads and shared memory Shared-memory communication gives rise to interference dependence

Definition

Statement t is interference-dependent on statement s, if t uses a value which is defined by statement s, and s and t may happen in parallel

Example

1 thread_1: 2 a = input + c; 3 b = input - 5; 4 thread_2: 5 c = b;

2

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-6
SLIDE 6

cSDG – concurrent SDG

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. 1 2 3 4 5

3

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-7
SLIDE 7

cSDG – concurrent SDG

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. 1 2 3 4 5

Two interference dependences may exclude each other

3

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-8
SLIDE 8

cSDG – concurrent SDG

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. 1 2 3 4 5

Two interference dependences may exclude each other ⇒ Time-insensitive slices

3

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-9
SLIDE 9

Time-sensitive Slicing

Time-sensitive path

A context-sensitive path n1, . . . , nk in a cSDG is time-sensitive, if ∀1 ≤ i < j ≤ k : ni and nj may happen in parallel, or ni reaches nj in the control flow graph

4

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-10
SLIDE 10

Time-sensitive Slicing

Time-sensitive path

A context-sensitive path n1, . . . , nk in a cSDG is time-sensitive, if ∀1 ≤ i < j ≤ k : ni and nj may happen in parallel, or ni reaches nj in the control flow graph

Prepending property (Krinke, 2003)

Let π = n1, . . . , nk be a time-sensitive path in a cSDG G. Let e = n0 → n1 be an edge in G. Path n0, n1, . . . , nk is time-sensitive, if e is thread-local, or n0 reaches the first element n in π that may not happen in parallel to n0

4

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-11
SLIDE 11

Time-sensitive Slicing

Exploit prepending property: Annotate nodes with state tuples → One entry per thread → Contains the first element of that thread in the path taken so far

5

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-12
SLIDE 12

Time-sensitive Slicing

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. [2,X] 1 2 3 4 5

6

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-13
SLIDE 13

Time-sensitive Slicing

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. [2,X] 1 2 3 4 5 [1,X]

6

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-14
SLIDE 14

Time-sensitive Slicing

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. [2,X] 1 2 3 4 5 [1,X] [2,5]

6

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-15
SLIDE 15

Time-sensitive Slicing

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. [2,X] 1 2 3 4 5 [1,X] [2,5] [2,4]

6

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-16
SLIDE 16

Time-sensitive Slicing

Runtime costs A node can be visited repeatedly with different state tuples State tuple entry is node + calling context ⇒ O(Npt), N = number of nodes, p = max. call depth, Np = upper bound for nodes + calling contexts, t = number of threads

7

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-17
SLIDE 17

Time-sensitive Slicing

Practical for programs with approx. 10 kLoc Usable for mature languages

Interprocedural programs, including recursion Dynamic thread creation inside loops or recursion

JOANA-Project for full Java bytecode (Giffhorn and Hammer, Precise Analysis of Java Programs using JOANA (Tool Demonstration), in 8th IEEE SCAM, 2008.

8

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-18
SLIDE 18

Empirical Results

name nodes edges methods threads

  • interf. dep.

DiningPhils 5143 125470 116 2 471 LaplaceGrid 6218 51035 161 2 948 Barcode 12393 64820 271 2 5 HyperM 17835 97827 277 6 8139 Podcast 23676 159478 407 3 128 1,000 slices per program: name size per slice (nodes) runtime per slice (msec) cont.-sens. time-sens. cont.-sens. time-sens. DiningPhils 2,867 2,499 (87%) 25 6,711 LaplaceGrid 3,409 3,328 (98%) 12 10,167 Barcode 3,410 2,974 (87%) 12 275 HyperM 9,222 7,441 (81%) 36 2,888 Podcast 12,335 8,730 (71%) 50 7,286

9

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-19
SLIDE 19

Important Topics for Future Research Runtime costs Prepending property and MHP information Interference dependence and reaching definitions

10

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-20
SLIDE 20

Runtime Costs

What property of a cSDG has the major influence on the runtime costs? name nodes edges methods threads

  • interf. dep.

DiningPhils 5143 125470 116 2 471 LaplaceGrid 6218 51035 161 2 948 Barcode 12393 64820 271 2 5 HyperM 17835 97827 277 6 8139 Podcast 23676 159478 407 3 128 name size per slice (nodes) runtime per slice (msec) cont.-sens. time-sens. cont.-sens. time-sens. DiningPhils 2,867 2,499 (87%) 25 6,711 LaplaceGrid 3,409 3,328 (98%) 12 10,167 Barcode 3,410 2,974 (87%) 12 275 HyperM 9,222 7,441 (81%) 36 2,888 Podcast 12,335 8,730 (71%) 50 7,286

11

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-21
SLIDE 21

Runtime Costs

What property of a cSDG has the major influence on the runtime costs? name nodes edges methods threads

  • interf. dep.

DiningPhils 5143 125470 116 2 471 LaplaceGrid 6218 51035 161 2 948 Barcode 12393 64820 271 2 5 HyperM 17835 97827 277 6 8139 Podcast 23676 159478 407 3 128 name size per slice (nodes) runtime per slice (msec) cont.-sens. time-sens. cont.-sens. time-sens. DiningPhils 2,867 2,499 (87%) 25 6,711 LaplaceGrid 3,409 3,328 (98%) 12 10,167 Barcode 3,410 2,974 (87%) 12 275 HyperM 9,222 7,441 (81%) 36 2,888 Podcast 12,335 8,730 (71%) 50 7,286

11

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-22
SLIDE 22

Runtime Costs

What property of a cSDG has the major influence on the runtime costs? name nodes edges methods threads

  • interf. dep.

DiningPhils 5143 125470 116 2 471 LaplaceGrid 6218 51035 161 2 948 Barcode 12393 64820 271 2 5 HyperM 17835 97827 277 6 8139 Podcast 23676 159478 407 3 128 name size per slice (nodes) runtime per slice (msec) cont.-sens. time-sens. cont.-sens. time-sens. DiningPhils 2,867 2,499 (87%) 25 6,711 LaplaceGrid 3,409 3,328 (98%) 12 10,167 Barcode 3,410 2,974 (87%) 12 275 HyperM 9,222 7,441 (81%) 36 2,888 Podcast 12,335 8,730 (71%) 50 7,286

11

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-23
SLIDE 23

Runtime Costs

What property of a cSDG has the major influence on the runtime costs? name nodes edges methods threads

  • interf. dep.

DiningPhils 5143 125470 116 2 471 LaplaceGrid 6218 51035 161 2 948 Barcode 12393 64820 271 2 5 HyperM 17835 97827 277 6 8139 Podcast 23676 159478 407 3 128 name size per slice (nodes) runtime per slice (msec) cont.-sens. time-sens. cont.-sens. time-sens. DiningPhils 2,867 2,499 (87%) 25 6,711 LaplaceGrid 3,409 3,328 (98%) 12 10,167 Barcode 3,410 2,974 (87%) 12 275 HyperM 9,222 7,441 (81%) 36 2,888 Podcast 12,335 8,730 (71%) 50 7,286

11

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-24
SLIDE 24

Prepending Property and MHP Information

May-happen-in-parallel analysis: Which parts of the threads may happen in parallel? Recall the definition of time-sensitive paths:

Time-sensitive path

A context-sensitive path n1, . . . , nk in a cSDG is time-sensitive, if ∀1 ≤ i < j ≤ k : ni and nj may happen in parallel, or ni reaches nj in the control flow graph Time-sensitivity depends on available MHP information In our evaluation it increased precision by 10%

12

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-25
SLIDE 25

Prepending Property and MHP Information

May-happen-in-parallel analysis: Which parts of the threads may happen in parallel? Recall the definition of time-sensitive paths:

Time-sensitive path

A context-sensitive path n1, . . . , nk in a cSDG is time-sensitive, if ∀1 ≤ i < j ≤ k : ni and nj may happen in parallel, or ni reaches nj in the control flow graph Time-sensitivity depends on available MHP information In our evaluation it increased precision by 10% Cannot be completely exploited by the time-sensitive slicer

12

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-26
SLIDE 26

Prepending Property and MHP Information

Prepending property (Krinke, 2003)

Let π = n1, . . . , nk be a time-sensitive path in a cSDG G. Let e = n0 → n1 be an edge in G. Path n0, n1, . . . , nk is time-sensitive, if e is thread-local, or n0 reaches the first element n in π that may not happen in parallel to n0 Does only hold in case all threads may happen in parallel ⇒ n0 and n may not happen in parallel to the same elements M in π ⇒ n reaches every m ∈ M ⇒ ‘Reaches’ is transitive, thus it suffices that n0 reaches n

13

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-27
SLIDE 27

Prepending Property and MHP Information

Example

1 thread_1: 2 if (...) 3 x = 1; 4 else 5 fork thread_2; 6 y = x; 7 thread_2: 8 print y;

14

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-28
SLIDE 28

Prepending Property and MHP Information

thread_1 x = 1 if (...) thread_2 print y fork thread_2 y = x 1 2 3 5 6 7 8 control dep. interference dep. data dep.

15

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-29
SLIDE 29

Prepending Property and MHP Information

thread_1 x = 1 if (...) thread_2 print y fork thread_2 y = x 1 2 3 5 6 7 8 [X,8] [5,8] [3,8] control dep. interference dep. data dep.

15

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-30
SLIDE 30

Prepending Property and MHP Information

Develop a slicer that is time-sensitive wrt. general MHP information Adjust the prepending property? ⇒ Which information has to be stored in the state tuples? Find a completely different approach?

Post-process time-insensitive slices Deactivate invalid parts of the cSDG (use Petri nets?) . . .

16

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-31
SLIDE 31

Interference Dependence and Reaching Definitions

Definition

Statement t is interference-dependent on statement s, if t uses a variable v which is defined by statement s, and s and t may happen in parallel Data dependence requires reaching definitions Interference dep. ignores reaching definitions Undecidable for concurrent interprocedural programs (Müller-Olm and Seidl, On optimal slicing of parallel programs, in ACM Symposium on Theory of Computing, 2001) Decidable in special cases

17

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-32
SLIDE 32

Threaded Interaction Reachability Graph

Idea of Qi et al. (Slicing Concurrent Programs Based on Program Reachability Graphs, in IEEE 10th ICQS, 2010.) Unroll the possible interleavings in a Threaded Interaction Reachability Graph (TIRG) Each node corresponds to a possible interleaving situation

18

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-33
SLIDE 33

Threaded Interaction Reachability Graph

thread_1 b = input - 5 a = input + c thread_2 c = b 1 2 3 4 5 [1,4] [2,4] [2,5] [3,5] [3,4] [1,5]

19

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-34
SLIDE 34

Threaded Interaction Reachability Graph

Create a SDG from the TIRG ⇒ Data flow analysis for sequential programs Free of interference dependence

No time-insensitivity Finds situations, where a def in one thread does not reach a use in another thread

20

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-35
SLIDE 35

Threaded Interaction Reachability Graph

Create a SDG from the TIRG ⇒ Data flow analysis for sequential programs Free of interference dependence

No time-insensitivity Finds situations, where a def in one thread does not reach a use in another thread

Very high precision No empirical data ⇒ assumption: TIRGs are huge (need to inline procedures) No recursion No thread creation inside loops ⇒ Extend its power ⇒ Develop compression techniques

20

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-36
SLIDE 36

Literature

  • J. Krinke, Context-sensitive Slicing of Concurrent Programs,

in 11th Foundations of Software Engineering, 2003

  • M. G. Nanda and S. Ramesh, Interprocedural Slicing of Multithreaded

Programs with Applications to Java, in ACM TOPLAS, 28(6):1088–1144, 2006.

  • D. Giffhorn and C. Hammer, Precise Slicing of Concurrent Programs -

An Evaluation of Static Slicing Algorithms for Concurrent Programs, in Springer JASE, 16(2):197–234, 2009.

21

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-37
SLIDE 37

JIT Compiler

Just-in-time (JIT) compiler may reorganize code during program execution Control flow during execution may not correspond to the control flow during the slice ⇒ Time-insensitive paths may turn time-sensitive

22

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-38
SLIDE 38

JIT Compiler

thread_1 b = input - 5 a = input + c thread_2 c = b 1 2 3 4 5 thread_1 a = input + c thread_2 c = b 1 2 3 4 5 b = input - 5

Java’s JIT compiler may switch statements 2 and 3

23

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting

slide-39
SLIDE 39

JIT Compiler

control dep. thread_1 b = input - 5 a = input + c thread_2 c = b interference dep. 1 2 3 4 5

Slice is time-sensitive! Identification of which execution orders are guaranteed

24

  • 19. Januar 2011

Dennis Giffhorn – Slicing Concurrent Programs Programming paradigms group – IPD Snelting