BIRS Workshop, Banff, Canada Jan 22, 2014 © 2014 IBM Corporation Resolution and Parallelizability:...

BIRS Workshop, Banff, Canada

Jan 22, 2014 © 2014 IBM Corporation

Resolution and Parallelizability:Barriers to the Efficient Parallelization of SAT Solvers

George Katsirelos MIAT, INRA, Toulouse, FranceAshish Sabharwal IBM Watson, USAHorst Samulowitz IBM Watson, USALaurent Simon Univ. Paris-Sud, LRI/CNRS, Orsay, France

[published at AAAI-2013]

Resolution and Parallelizability

© 2014 IBM Corporation

Trend Towards Parallelization

Focus Shifting From Single-Thread Performanceto Multi-Processor Performance

– 100s and even 1000s of compute cores easily accessible

– Classical Algorithm Parallelization, e.g., parallel sort, shortest path,PRAM model, AC circuits

– Significant Advances in Data Parallelisme.g., MapReduce, Hadoop, SystemML, R statistics

Challenge: Search and Optimization on 1000s of Processors

– Tremendous advances in the Sequential case of Combinatorial Search E.g., SAT solvers can tackle instances with ~2M variables, 10M constraints!

– Exponential search appears to be an “obvious” candidate to parallelize!

– In fact, many SAT/CSP/MIP solvers already do support multi-core andmulti-machine runs

2 Banff Workshop on SAT, 2014 | Katsirelos, Samulowitz, Sabharwal, Simon



Parallelization of Combinatorial Search

Fact: State-of-the-Art Search Engines Do NOT Parallelize Well

– Brute Force exponential search is, of course, trivial to parallelize

– But sophisticated search engines that adapt (through e.g. clause learning, variable impact aggregation, etc.) have inherent sequential aspects

– Modern SAT/MIP/”adapting”-CP solvers do not parallelize well• Supporting data: next slide

AAAI 2012 Challenge Paper on the topic [Hamadi & Wintersteiger 2012]

– P-completeness of Unit Propagation a key barrier (solvers spend ~80% of the time Unit Propagating and we don’t know how to parallelize P well)

– Our result: barriers exist even if Unit Propagation came for free!




Parallelization of Combinatorial Search: SAT

Rather Disappointing Performance at SAT Competitions – e.g., in 2011:

– Average speedup on 8 cores only ~1.8x, on 32 cores only ~3x

– Top performing parallel solvers were based on little to no communication(CryptoMinisat-MT [Soos 2012], Plingeling [Biere 2012])

– Winners were “simple” Portfolio solvers (ppfolio [Roussel], pfolioUZK [Wotzlaw et al])

Plingeling-ats-587[Dec 2013]

– Single machine with 128 coresand 128 GB memory

– Benchmark set used in thiswork, restricted to the 142instances solved by 1 core in[10,5000) seconds


1 6 640.50

5.00

1.00

1.25

1.90

2.57 2.69

1.63

Plingeling ats 587

Number of Cores

Sp

ee

du

p(g

eo

me

tric

av

era

ge

)


© 2014 IBM Corporation5 Banff Workshop on SAT, 2014 | Katsirelos, Samulowitz, Sabharwal, Simon

What makes parallelization of SAT solvers hard?

Can we obtain insights into their behaviorbeyond eventual wall-clock performance?



Contributions of the Work

A New Systematic Study of Parallelism in the Context of Searchthrough the Lens of Proof Complexity

– Focus on understanding rather than on engineering

– Are there inherent bottlenecks that may hinder parallelization,irrespective of which heuristics are used to share information?

1. A Practical Study: Interesting properties of Actual Proofs

– Proofs generated by state-of-the-art SAT solvers contain narrow bottlenecks

2. Proof-Based Measures that capture Best-Case Parallelizability

– Coarse measure: “Depth” of the proof graph

– Refined measure: Makespan of a resource constrained scheduling problem

3. Empirical Findings: Correlations and Parallelization Limits

– Typical sequential proofs are not very parallelizable even in the best case!

– “Schedule speedup” / makespan correlates with observed speedup




Approach: Proof Complexity (applied here to Typically Generated Proofs)

Proof Complexity [Cook & Reckhov, 1979]: Study of the nature (e.g., size, width, space, depth, “shape”, etc.) of Proofs of Unsatisfiability

– Resolution Graph of Conflict-Directed-Clause-Learning (CDCL) SAT Solvers

Runtime(any SAT solver, F) minproofs Size(Resolution proof of F)

– Note: Insights applicable also to Satisfiable instances!• Solvers prove a lot of sub-formulas to be unsatisfiable before hitting the first solution• Formal characterization [Achlioptas et al, 2001 & 2004]

Study of Proofs has provided strong insights into CDCL SAT solvers

– What does “clause learning” bring?

– What do “restarts” add?

[Beame et al, 2004; Buss et al, 2008, 2012; Hertel et al, 2008; Pipatsrisawat et al, 2011]


Worst case / Best case results


© 2014 IBM Corporation8

Underlying Inference Principle: Resolution

CDCL SAT solvers produce Resolution Derivations

Proof Graph and Depth:

– Each initial and derived constraint is a node, annotated with its proof depth

– proofdepth(initial clause C) = 0

– proofdepth(derived clause C) = 1 + maxparents proofdepth(parent(C))

C1 0 C2 0 C3 0 C4 0 C5 0 C6 0

C7 1

C8 2

C9 1

C10 3

C11 2

C12 3

C13 4

Banff Workshop on SAT, 2014 | Katsirelos, Samulowitz, Sabharwal, Simon

Constraint ID Depth

F :



How Parallelizable are Resolution Refutations?

Refutation(F) = Resolution Proof that derives the empty (“false”) clause

Depth of the proof clearly limits the amount of potential parallelization

– Chain of dependencies

– Theorem: All Resolution Proof Graphs of certain “pebbling” style instances have large depth; also holds for all Conflict Resolution Graphs (XOR substitution trick)

However, proofdepth bound on parallelization is very crude

– Does not explain poor performance with small k (e.g., 8, 32, … processors)

How does a typical sequential SAT solver proof look like?

– Setup for Experiments:• Sequential Glucose 2.1 extended with proof output• GluSatX10: using SatX10 to run a k-processor version of Sequential Glucose

– Working Assumption: Proofs produced by GluSatX10 on k cores look “similar”to proofs produced by Sequential Glucose


http://x10-lang.org/satx10 [IBM Teams: X10 and SAT/CSP]

** simplified statements; see paper for more formal notions



Proof Graph Example: Very Complex Structure


[Easy sequential case, solved in ~30 seconds]



Bottlenecks in Typical SAT Proofs


Proofs Generated by SAT Solvers Exhibit Surprisingly Narrow “Bottlenecks”, i.e., Depths with Very Few (~1) Clauses!

– Nothing deeper can be derived before bottleneck clauses Sequentiality

Depth in the proof

Nu

mb

er o

f C

lau

ses

(lo

g-sc

ale)

Der

ived

at

that

Dep

th



Best-Case Parallelization with k Processors

Given Proof P and k Processors, Best-Case Parallelization of P = Resource Constrained Scheduling Problem with Precedences

Let Mk(P) = makespan of the optimal schedule of P on k processors

– Even approximating Mk(P) within 4/3 is NP-hard, but (2 – 1/k) approx. is easy

Best-Case k processor speedup on P: Sk(P) = M1(P) / Mk(P)


C1 0 C2 0 C3 0 C4 0 C5 0 C6 0

C7 1

C8 2

C9 1

C10 3

C11 2

C12 3

C13 4Constraint ID Depth

C’9 1Example:M1(P) = 8M2(P) = 5M3(P) = 4M4(P) = 4…depth = 4

1 1 2

2 3

3 4

5



Makespan vs. Proof Depth

Schedule Makespan yields a finer grained lower bound, Sk(P),on best-case parallelization than proof depth

– proofdepth(P) : limit of parallelization of P with “infinite” processors

– Mk(P) proofdepth(P)

– Mk(P) proofdepth(P) as k




Even Best-Case Parallelization Efficiency is Low Beyond 100 Processors


Best-Case Efficiency of parallelizing P with k processors = 100 * (Sk(P) / k)

E.g., 100% = full utilization of k processors speedup = k



Proofs of Some Instances Exhibit Very LowBest-Case Schedule Speedup


A) Even with 1024 processors,best-case speedup ~ 50-100

B) 128 processors insufficient toachieve a speedup of ~ 90



Best-Case Schedule Speedup Correlates WithActual Observed Runtime Speedup


Average over a sliding window

(Makes the study of the best-case schedule speedup relevant)



Summary

A New Systematic Study of Parallelism in the Context of Searchthrough the Lens of Proof Complexity

– Focus on understanding rather than on engineering

Main Findings:

A. Typical Sequential Refutations Contain Surprisingly Narrow Bottlenecks

B. Typical Sequential Refutations are Not Parallelizable Beyond a Few Processors, even in the best case of offline ‘schedule speedup’ produced in hindsight

C. Observed Runtime Speedup with k processors weakly correlates withBest-Case Schedule Speedup of a Sequential Proof produced in hindsight

Open Question: Can we design SAT solvers that generate Proofs that are inherently More Parallelizable?

Caveat: assumption that proofs generated by GluSatX10 on k cores look “similar” to proofs generated by Sequential Glucose


BIRS Workshop, Banff, Canada Jan 22, 2014 © 2014 IBM Corporation Resolution and Parallelizability:...

Documents

Transcript of BIRS Workshop, Banff, Canada Jan 22, 2014 © 2014 IBM Corporation Resolution and Parallelizability:...