SAT Based Efficient Directed Test Generation Techniques

Transcription

1 SAT Based Efficient Directed Test Generation Techniques Presented by Mingsong Chen Software Engineering Institute East China Normal University May 5, 2011

2 Outline Introduction Model Checking Based Test Generation SAT-based Bounded Model Checking DPLL algorithm Conflict clause Efficient Test Generation Approaches Conflict clause forwarding based approaches Decision ordering based techniques Property decomposition based methods Conclusion 2

3 Outline Introduction Model Checking Based Test Generation SAT-based Bounded Model Checking DPLL algorithm Conflict clause Efficient Test Generation Approaches Conflict clause forwarding based approaches Decision ordering based techniques Property decomposition based methods Conclusion 3

4 Functional Validation of SOC Designs Engineer Years B 10B 100M Simulation Vectors 1M 10M 100M Logic Gates Source: G. Spirakis, keynote address at DATE 2004 Source: Synopsys Functional validation is a major bottleneck during SoC development! (up to 70% of time and resources are used)

5 Functional Validation Methods Simulation (Validation) The process of gaining confidence by examining the behavior of the implementation using input/output test vectors Incompleteness verification: not possible for all input vectors Applicable to large designs Formal (Verification) Mathematical proof that a system (implementation) behaves according to a given set of requirements (specification) Complete verification Applied to small and critical components due to the state space explosion problem 5

6 Approaches for Specification Validation Verification / Validation Formal Verification Simulation-based Validation Theorem Proving Model Checking SAT Solving Random Test Constrained Random Test Directed Test Bounded Model Checking SAT-based Bounded Model Checking Validation using a combination of simulation based techniques and formal methods. 6

7 Test Generation using Model Checking Model Checking (MC) Specification is translated to formal models, e.g., SMV Desired behaviors in temporal logic properties, e.g. LTL Property falsification leads to counterexamples (tests) Test Generation Generate a counterexample: sequence of variable assignments ATM Model Password Input is always true Problem: Test generation is very costly or not applicable in many complex scenarios. An Example Model Checker Approach: Generate a Exploit test to fail passward learning input to reduce validation complexity User name Password User intput - Reduction of test generation time Bob ABC ABD - Enables test generation in complex scenarios 7

8 Outline Introduction Model Checking Based Test Generation SAT-based Bounded Model Checking DPLL algorithm Conflict clause Efficient Test Generation Approaches Conflict clause forwarding based approaches Decision ordering based techniques Property decomposition based methods Conclusion 8

9 SAT-based Bounded Model Checking The safety property P is valid up to cycle k iff Ω(k) is not satisfiable. p p p p p... s 0 s 1 s 2 s k-1 s k If Ω(k) is satisfiable, then we can get an assignment which can be translated to a test.

10 SAT Decision Procedure Given a ϕ in CNF: (x+y+z)( x+y)( y+z)( x+ y+ z) x ϕ ϕ x Decide() (y),( y,z),( y, z) (y,z),( y,z) x=1@ level1 y y z z (z),( z) () () (y),( y) z=1@level2 z z X y y () () () () X X X X Deduce() Resolve_Conflict() y=0 or 1@level2

11 DPLL Algorithm while (1){ run_periodic_function(); if( decide_next_branch() ){ while (deduce() Implication == CONFLICT) { } blevel = analyze_conflicts(); Conflict Backtrack if( blevel<0 ) return UNSAT; } } else return SAT; BCP = Implication Number + Conflict Backtrack Boolean Constraint Propagation (BCP) consumes up to 80% of the time and resources during SAT solving

12 Implication Graph, Conflict Clause ω 1 = (x 2 x 6 x 4 ) ω 2 = ( x 8 x 3 x 7 ) ω 3 = ( x 1 x 4 x 5 ) ω 4 = ( x 3 x 4 ) ω 5 = ( x 2 x 3 x 8 ) ω 6 : ( x 1 x 5 x 6 x 7 ) x5@4 x1@3 x4@4 x6@1 x2@4 x3@4 cut1 x8@4 x8@4 conflict x7@2 Conflict clause can be treated as the knowledge learned during the SAT solving. It is a restriction of the variable assignment.

13 Same Property but Different Bounds p 1 k p 1 3 p 1 2 p 1 1 Δp 1 k p 1 3 Forward Δp 1 3 Forward p 1 1 p 1 2 Δp 1 2 The minimal bound is k: Save: ΔP Δp Δp 1 k Δp 1 k O. Strichman. Pruning Techniques for the SAT-Based Bounded Model Checking Problems. CHARME, 2001

14 Outline Introduction Model Checking Based Test Generation SAT-based Bound Model Checking DPLL algorithm Conflict clause Efficient Test Generation Approaches Conflict clause forwarding based approaches Decision ordering based techniques Property decomposition based methods Conclusion 14

15 Same Design, Different Properties P1 P2 rg2 rg3 rg1 rgk rg4 rb2 P3 rb1 rb4 rb3 rbn Benefit: Original: Red + Blue + Green Now: Red + (Blue Δblue) + (Green Δgreen) Save: Δblue + Δgreen Forward Forward Δgreen Δblue

16 Property Clustering Clustering properties is to exploit the structural and behavior similarity and maximize the validation reuse Property clustering methods: Based on structural similarity Based on textual similarity Based on Influence (Cone of Influence) Based on CNF intersections

17 Identification of Common Conflict Clauses cut1 conflict Conflict Clause ( X1 X5 X6 X7 ) Conflict Side Clauses Clauses Group ID ( X2 X3 X8 ) (X3 X7 X8 ) (X2 X3 X6 ) ( X3 X4) ( X1 X4 X5 ) Let be the bit AND operation. ( ) = So the conflict clause ( X1 X5 X6 X7 ) can be reused for property 2.

18 Test Generation For A Property Cluster 1. Cluster the properties based on similarity 2. for each cluster i, of properties 1 Select base property p i 1, and generate CNF i 1 2 for each CNF i j of p i j (j 1) in cluster i a) Perform name substitution on CNF i j b) Compute intersection INT i j between CNF i 1 and CNF i j c) Mark the clauses of CNF i 1 using INT i j endfor 3 Solve CNF i 1 to get the conflict clauses CC i 1 and test i 1 4 for each CNF i j (j 1) a) CNF i j = CNF i j + Filter (CC i j, j) b) Solve CNF i j to get test i j endfor endfor

19 Case Study 1 : MIPS Processor PC Fetch Memory Register File Decode The Architecture IALU MUL1 MUL2 FADD1 FADD2 FADD3 DIV MIPS Processor - 20 nodes - 24 edges - 91 instructions MUL7 FADD4 Decode WriteBack Unit Storage Pipeline edge Data-transfer edge

20 MIPS Processor Results The processor has five pipeline stages: fetch, decode, execute, memory and writeback. There are totally 171 properties generated. Methods Structure Textual Influence Intersection Num. of Clusters zchaff (sec.) (Existing Approach) Our Method (sec.) Speedup zchaff is a state-of-the-art SAT Solver.

21 Case Study 2 : OSES VerifyOrd er Settle_tra de Trade_S Update_SHolderDB _S Update_StockDB _S Update_OrderDB_ S Update_orderDB_ NM Trade_N M Trade_ PE Update _stockdb_pe Update_StockerHolderDB_ PE Update_OrderDB_ PE UpdateM ap AddOrderForm List GetNewOrd er Trade _F CheckLimitPr ice Update_orderDB _F End Order Order Error Get Order Result Limit Buy Market Buy Marker Sale Limit Sale t0 t11 t1 t2 t12 t9 t8 t6 t5 t7 t3 t15 t16 t13 t14 t17 t18 t10 t22 t27 t28 t29 t23 t29 t26 t2 5 t2 4 t20 t21 t19 t4

22 OSES Results This case study is a on-line stock exchange system. The activity diagram consists of 27 activities, 29 transitions and 18 key paths. There are totally 51 properties. Methods Structure Textual Influence Intersection Num. of Clusters zchaff (sec.) (Existing Approach) Our Method (sec.) Speedup

23 Outline Introduction Model Checking Based Test Generation SAT-based Bounded Model Checking DPLL algorithm Conflict clause Efficient Test Generation Approaches Conflict clause forwarding based approaches Decision ordering based techniques Property decomposition based methods Conclusion 23

24 Decision Ordering Problem Given a ϕ in CNF: (x+y+z)( x+y)( y+z)( x+ y+ z) ϕ ϕ x x (y),( y,z),( y, z) (y,z),( y,z) y y z z (z),( z) () () (y),( y) z z X y y A wise decision ordering can quickly locate the true assignment. Bit value ordering Variable Orderinig () () () () X X X X Best decision: x, z 24

25 Two Similar SAT Problems SAT 1 SAT 2 a a b b b b c c c c c c c c F F F S F F F F F F F F F F F S Ordering: a, a, b, b, c, c Ordering: a, a, b, b, c, c Without Learning, 7 conflicts in SAT2.

26 Learning: Bit Value Ordering SAT 1 SAT 2 a Bit value: a=1,b=0,c=0 a b b b b c c c c c c c c F F F S F F F F F F F F F F F S Ordering: a, a, b, b, c, c Ordering: a, a, b, b, c, c With bit value learning, 4 conflicts in SAT2.

27 Learning: Bit Value + Variable Ordering SAT 1 SAT 2 Bit value: a=1,b=0,c=0 a Variable order: b>c>a b b b c c c c c c F F F S F F F F Ordering: a, a, b, b, c, c a a a a F F F F F F F S Ordering: b, b, c, c, a, a With bit value+ variable order learning, 1 conflict in SAT2.

28 Our method An Example with 3 properties VarStat a b c d [0] V [1] V P1: a=0, b=0, c=1, d=1 VarStat a b c d [0] V [1] V P2: a=0, b=0, c=1, d=0 P3: a=0, b=0, c=1, d=? score(a), score(a ) score(b), score(b ) score(c), score(c ) Predict ordering for P3 VarStat a b c d [0] V [1] V Approach: Using the statistics of the counterexamples when checking the properties in a cluster - Count of values bit value ordering - Variance of counts of two literals variable ordering

29 Case Study 1 : MIPS Processor For each function unit (ALU, DIV, FADD and MUL) in the pipelined processor. We generate 4 properties. Property (test) zchaff (sec) Clustering Speedup (over zchaff) Decision Ordering Speedup (over Clustering) ALU P P P DIV P P P FADD P P P MUL P P P

30 Case Study 1 : MIPS Processor Test generation time is significantly improved - Drastic reduction of conflict clauses - Drastic reduction in number of implications

31 Case Study 2 : OSES This case study is a on-line stock exchange system. The activity diagram consists of 27 activities, 29 transitions and 18 key paths. Cluster Size zchaff Clustering Speedup (over zchaff) Decision Ordering Speedup (over Clustering) C C C C C C C C C Average

32 Outline Introduction Model Checking Based Test Generation SAT-based Bounded Model Checking Implication graph SAT decision procedure DPLL algorithm Efficient Test Generation Approaches Conflict clause forwarding based approaches Decision ordering based techniques Property decomposition based methods Conclusion 32

33 Property Decomposition Techniques Property Property p1 p2 pn p1 p2 pn t1 t2 tn Learnings Drawback: Composition Hard to automate BMC Test Koo et al. Functional Test Generation using Property Decomposition Techniques. ACM TECS, 2009 Test 33

34 Spatial Decomposition V1 V2 Cone1 p1 COI(p1) < COI(p2) < COI(p3) <COI(P) V3 V4 V5 Cone2 p2 P Time(p1) < Time(p2) < Time(p3) <Time(P) Cone3 p3 Vn Learning from P1 can reduce the Time(P)? 34

35 Temporal Decomposition T2 T1 T3 e1 e3 e4 e5 e2 e6 Cause effect relation: e1 e2 e3 e4 e5 e6 Happen before relation: e1<e3<e4 <e5<e2<e6 35

36 Temporal Decomposition e3 e4 e5 1 2 e1 1 e2 3 e7 e8 e e6 5 event Cause-effect Happen-before!F(e1)!F(e3)!F(e7)!F(e9) 36

37 Case Study 1: MIPS Processor We generated 6 complex properties based on interaction faults on various function unit (ALU, DIV, FADD and MUL), which cannot handled by temporal decomposition. Property (test) zchaff (sec) Num. of Clusters Num. of Sub-props Spatial (sec) Speedup P P P P P P Speedup: times

38 Case Study 2 : OSES This case study is a on-line stock exchange system. The activity diagram consists of 27 activities, 29 transitions and 18 key paths. Property zchaff (sec) Bound Num. of Subproperties Temporal (sec) Speedup P P P P P P P P P P Speedup: 3-62 times

39 Outline Introduction Model Checking Based Test Generation SAT-based Bounded Model Checking DPLL algorithm Conflict clause Efficient Test Generation Approaches Conflict clause forwarding based approaches Decision ordering based techniques Property decomposition based methods Conclusion 39

40 Conclusion Validation is a major bottleneck in HW/SW designs This presentation discusses how to reduce the overall validation effort for directed test generation from models. 1. Conflict clause forwarding and property clustering methods 2. Efficient decision ordering approaches 3. Property decomposition techniques Successfully applied on both HW/SW designs Several orders of magnitude reduction in overall validation effort 40

41 Thank you!