parallel qbf solving state of the art techniques and
play

Parallel QBF Solving: State of the Art Techniques and Future - PowerPoint PPT Presentation

Parallel QBF Solving: State of the Art Techniques and Future Perspectives Florian Lonsing Knowledge-Based Systems Group, Vienna University of Technology, Austria http://www.kr.tuwien.ac.at/staff/lonsing/ First Workshop on Parallel Constraint


  1. Parallel QBF Solving: State of the Art Techniques and Future Perspectives Florian Lonsing Knowledge-Based Systems Group, Vienna University of Technology, Austria http://www.kr.tuwien.ac.at/staff/lonsing/ First Workshop on Parallel Constraint Reasoning (PCR’17) August 6, 2017, Gothenburg, Sweden Collocated with CADE-26 This work is supported by the Austrian Science Fund (FWF) under grant S11409-N23. Florian Lonsing (TU Wien) Parallel QBF Solving 1 / 15

  2. Introduction Quantified Boolean Formulas (QBF): Existential ( ∃ ) / universal ( ∀ ) quantification of propositional variables. Checking QBF satisfiability: PSPACE-complete. Potentially more succinct encodings than propositional logic. Example QBF ψ := ˆ Q .φ in prenex conjunctive normal form (PCNF) . ψ = ∀ u ∃ x . (¯ u ∨ x ) ∧ ( u ∨ ¯ x ) . � �� � � �� � quantifier prefix propositional CNF Recursive semantics: Assign variables in prefix ordering, recurse on simplified formula ψ [ A ] under assignment A . Base cases: ⊥ is unsatisfiable, ⊤ is satisfiable. ∀ u .ψ is satisfiable iff ψ [ u / ⊥ ] and ψ [ u / ⊤ ] are satisfiable. ∃ x .ψ is satisfiable iff ψ [ x / ⊥ ] or ψ [ x / ⊤ ] is satisfiable. Florian Lonsing (TU Wien) Parallel QBF Solving 1 / 15

  3. Introduction: Sequential QBF Solving QCDCL: [GNT06, Let02, ZM02] Extension of conflict-driven clause learning (CDCL) for SAT solving. Derivation of new learned clauses and cubes (conjunctions of literals). Underlying proof system: Q-resolution calculus [KBKF95]. Expansion: [AB02, Bie04, JKMSC16, JM15b, RT15] Successively eliminate variables from a QBF. Shannon expansion [Sha49]: ∃ x .φ ( x , . . . ) ≡ φ ( ⊥ , . . . ) ∨ φ ( ⊤ , . . . ) Counter example guided abstraction refinement (CEGAR). QBF Proof Complexity: Recent results: orthogonality of proof systems [BCJ15, JM15a]. Expansion vs. Q-resolution: exponential separation wrt. proof sizes. Florian Lonsing (TU Wien) Parallel QBF Solving 2 / 15

  4. Outline State of the Art Techniques: Two main parallel approaches, inspired by parallel SAT solving. Examples of solvers illustrating the main approaches. Challenges: Limitations of parallel solving due to prefix ordering. Proof generation. Future Perspectives: Proof complexity (orthogonality) as a motivation to parallelize. Experiments to highlight performance diversity of sequential solvers. Florian Lonsing (TU Wien) Parallel QBF Solving 3 / 15

  5. Main Parallel QBF Solving Approaches Search Space Handling: Given a QBF with n variables, 2 n possible assignments. Additionally, prefix ordering and quantifier types must be considered (cf. semantics). Approaches differ in how search space is explored. Generalizations of approaches to parallel SAT solving. Example ψ = ∀ u ∃ x . (¯ u ∨ x ) ∧ ( u ∨ ¯ x ) . Out of the 2 2 possible assignments, { u , x } and {¬ u , ¬ x } are sufficient to show the satisfiability of ψ . Florian Lonsing (TU Wien) Parallel QBF Solving 4 / 15

  6. Main Parallel QBF Solving Approaches Portfolio Approach: Run several instances of the same solver (or different ones) on the original formula. Problem: instances potentially explore same parts of the search space. Diversify solver instances by parameter settings. With(out) sharing of learned clauses and cubes. Florian Lonsing (TU Wien) Parallel QBF Solving 4 / 15

  7. Main Parallel QBF Solving Approaches Splitting Approach: Instances by construction explore different parts of the search space. Parts are described by assignments that follow the prefix ordering. Instance gets original formula ψ and assignment A , and solves ψ [ A ] . Similar to guiding path method in parallel SAT solving. Example ψ = ∀ u ∃ x . (¯ u ∨ x ) ∧ ( u ∨ ¯ x ) . Possible splittings, starting from left end of prefix: A = { u } , solver instance works on ψ [ A ] = ∃ x . ( x ) . A = {¬ u } , solver instance works on ψ [ A ] = ∃ x . ( ¬ x ) . Unsound splittings: A = { x } , solver instance works on ψ [ A ] = ∀ u . ( u ) . A = {¬ x } , solver instance works on ψ [ A ] = ∀ u . ( ¬ u ) . Florian Lonsing (TU Wien) Parallel QBF Solving 4 / 15

  8. Portfolio Approach Example: HordeQBF MASTER 1 MASTER 2 MPI learned learned constraint constraint pool pool SM SM SM SM T 1 , 1 T 1 , 2 T 2 , 1 T 2 , 2 SM SM Massively parallel portfolio to run on distributed architectures [BL16]. First published parallel portfolio (cf. sequential ones [PT09, SM07]). Extension of HordeSat [BSS15], supports any QCDCL solver. Hierarchical parallelism: shared memory and message passing. Master executes identical but diversified QCDCL solvers (threads). Master processes: exchange of learned constraints via MPI. Florian Lonsing (TU Wien) Parallel QBF Solving 5 / 15

  9. Splitting Approach Example: MPIDepQBF M A n A n − 1 A i A 2 A 1 W 1 W 2 W n − 1 W n . . . Master M : Splits search space by assignments A i , calling workers W i . Combines results obtained by workers, further splitting. Manages exchange of learned constraints (not part of MPIDepQBF). Solvers differ in splitting strategy, exchange of learned constraints, shared memory vs. message passing [FMS00, LSB09, LSB + 11]. Florian Lonsing (TU Wien) Parallel QBF Solving 6 / 15

  10. Splitting Approach Example: MPIDepQBF M r n r n − 1 r i r 2 r 1 W 1 W 2 W n − 1 W n . . . Workers: Operate on original formula ψ , receive assignments A i and timeout. Treat A i as assumptions in QCDCL to solve ψ [ A i ] . Solving ψ under assumptions A i : re-use of learned constraints within each W i in next run of QCDCL under different A ′ i . Send result r i back to master or request increased timeout. Exchange of learned constraints either via master or among workers. Florian Lonsing (TU Wien) Parallel QBF Solving 6 / 15

  11. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 ∃ x 2 ∃ x 2 ( o , t 1 ) ( o , t 2 ) ( o , t 3 ) ( o , t 4 ) Initially 4 idle workers, 4 open leaves (subcases) with individual timeouts t i . Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  12. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 ∃ x 2 ∃ x 2 ( o , t 1 ) ( o , t 2 ) ( o , t 3 ) ( o , t 4 ) W 1 W 2 W 3 W 4 Assign open subcases to idle workers W i by sending assumptions: W 1 works on ψ [ ¬ x 1 , ¬ x 2 ] . W 2 works on ψ [ ¬ x 1 , x 2 ] . W 3 works on ψ [ x 1 , ¬ x 2 ] . W 4 works on ψ [ x 1 , x 2 ] . Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  13. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 ∃ x 2 ∃ x 2 u ( o , t 2 ) ( o , t 3 ) ( o , t 4 ) W 2 W 3 W 4 W 1 returns “unsat” for subcase ψ [ ¬ x 1 , ¬ x 2 ] and becomes idle. Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  14. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 ∃ x 2 ∃ x 2 u u ( o , t 3 ) ( o , t 4 ) W 3 W 4 W 2 returns “unsat” for subcase ψ [ ¬ x 1 , x 2 ] and becomes idle. Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  15. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u ∃ x 2 u u ( o , t 3 ) ( o , t 4 ) W 3 W 4 Since ψ [ ¬ x 1 , ¬ x 2 ] and ψ [ ¬ x 1 , x 2 ] unsatisfiable, also ψ [ ¬ x 1 ] unsatisfiable. Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  16. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u ∃ x 2 u u ( o , t 3 ) ( o , t 4 ) W 4 W 3 times out, W 1 , W 2 are idle, only 2 open leaves: generate new subcases. Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  17. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u ∃ x 2 u u ∀ y 3 ( o , t 4 ) W 4 ∃ x 4 ∃ x 4 ( o , t 5 ) ( o , t 6 ) ( o , t 7 ) ( o , t 8 ) Replace open leaf ( o , t 3 ) by binary tree based on ∀ y 3 and ∃ x 4 . Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  18. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u ∃ x 2 ( o , t 4 ) u u ∀ y 3 W 4 ∃ x 4 ∃ x 4 ( o , t 5 ) ( o , t 6 ) ( o , t 7 ) ( o , t 8 ) W 1 W 2 W 3 Assign open subcases to idle workers W 1 , W 2 , and W 3 by assumptions: W 1 works on ψ [ x 1 , ¬ x 2 , ¬ y 3 , ¬ x 4 ] . W 2 works on ψ [ x 1 , ¬ x 2 , ¬ y 3 , x 4 ] . W 3 works on ψ [ x 1 , ¬ x 2 , y 3 , ¬ x 4 ] . Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  19. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u ∃ x 2 ( o , t 4 ) u u ∀ y 3 W 4 ∃ x 4 ∃ x 4 ( o , t 6 ) ( o , t 8 ) s s W 2 W 1 and W 3 return “sat” for ψ [ x 1 , ¬ x 2 , ¬ y 3 , ¬ x 4 ] and ψ [ x 1 , ¬ x 2 , y 3 , ¬ x 4 ] . Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  20. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u ∃ x 2 ( o , t 4 ) u u ∀ y 3 W 4 s s ( o , t 6 ) ( o , t 8 ) s s W 2 Subcases ψ [ x 1 , ¬ x 2 , ¬ y 3 ] and ψ [ x 1 , ¬ x 2 , y 3 ] are satisfiable. Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  21. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u ∃ x 2 ( o , t 4 ) u u s W 4 s s ( o , t 6 ) ( o , t 8 ) s s W 2 Subcase ψ [ x 1 , ¬ x 2 ] is satisfiable. Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

  22. Splitting Approach Example PCNF ψ := ∃ x 1 , x 2 ∀ y 3 ∃ x 4 . . . φ . ∃ x 1 u s ( o , t 4 ) u u s W 4 s s ( o , t 6 ) ( o , t 8 ) s s W 2 Subcase ψ [ x 1 ] is satisfiable. Florian Lonsing (TU Wien) Parallel QBF Solving 7 / 15

Recommend


More recommend