FrequentSubtree MiningontheAutomataProcessor:ChallengesandOpportunities
ElahehSadredini,RezaRahimi,Ke Wang,KevinSkadronDepartmentofComputerScience
UniversityofVirginia
Presentingin:InternationalConferenceonSupercomputing,June13-16,2017
Motivation
11/21/17 InternationalConferenceonSupercomputing2017 2
Motivation
11/21/17 InternationalConferenceonSupercomputing2017 3
Motivation
11/21/17 InternationalConferenceonSupercomputing2017 4
WhatIstheFrequentSubtree Mining(FTM)?Ø To efficiently enumerate all frequent subtrees in a forest (database of trees) according to a given
minimum supportØ The support of a subtree is the number of subtrees in D that contains one occurrence of SØ A subtree S is frequent if its support is more than or equal to a user specified minimum support
value
11/21/17 InternationalConferenceonSupercomputing2017 5
AC
A
B C
D B
A
BA
DB
C
B C
AD
Relative support threshold: 60%
WhatIstheFrequentSubtree Mining(FTM)?Ø To efficiently enumerate all frequent subtrees in a forest (database of trees) according to a given
minimum supportØ The support of a subtree is the number of subtrees in D that contains one occurrence of SØ A subtree S is frequent if its support is more than or equal to a user specified minimum support
value
11/21/17 InternationalConferenceonSupercomputing2017 6
AC
A
B C
D B
A
BA
CB
C
B C
AD
Relative support threshold: 60%
Support = 3
Is frequent:
Preliminaries
Inducedsubtree
11/21/17 InternationalConferenceonSupercomputing2017 7
Embeddedsubtree
B
A
C
B
A C
C
B
A
C
Tree
Subtree
0
1
2
3
6
5
4
B
A
C
B
A C
C
B
A
C
Tree
Subtree
0
1
2
3
6
5
4
IssueswiththeCurrentFTMSolutions(1)Pros Cons
BFS Massive pruningMemory efficient
Multi-pass of datasetslow
DFS Fast Little pruning opportunityMemory-hungry
11/21/17 InternationalConferenceonSupercomputing2017 8
BFSandDFSrefertocandidategenerationapproach,nottreetraversalJ
IssueswiththeCurrentFTMSolutions(2)
11/21/17 InternationalConferenceonSupercomputing2017 9
IssueswiththeCurrentFTMSolutions(2)
11/21/17 InternationalConferenceonSupercomputing2017 10
Wouldthatbeacceptabletoachievehundreds-X speedup attheexpenseofloosingacoupleofpercentaccuracy?
Contributionofthisresearch:
• Proposing a memory efficient and fast solution to the frequentsubtree mining problem on the Automata Processor
• Achieving 350X and more speed up, when allowing 7.5% falsepositive subtrees
11/21/17 InternationalConferenceonSupercomputing2017 11
TheRestofThisTalk• AutomataProcessor• FTMChallengesandOpportunitiesontheAP• ExperimentalEvaluation• Takeaways
11/21/17 InternationalConferenceonSupercomputing2017 12
TheAutomataProcessor(1)• TheMicronAutomataProcessor(AP)isareconfigurablenon-vonNeumann
architecture,whichimplementsnon-deterministicfiniteautomata(NFA)withBooleanlogicgatesandcountersinhardwarebasedonDRAMtechnology.
RowAdd
ress
(Inpu
tSym
bol)
RowAccessresultsin49,152match&routeoperations(thenBooleanANDwith“active”bit-vector)
RoutingMatrix
Automata Processor
11/21/17 InternationalConferenceonSupercomputing2017 13
TheAutomataProcessor(2)• Amassivelyparallel‘MISD’architecture• 1Gbps dataprocessing• Hardwareresourcesondevelopmentboard
– StateTransitionElements(STE):1.5M– ReportingSTEs:200K– CounterElements:25K– BooleanElements:74K
11/21/17 InternationalConferenceonSupercomputing2017 14
ApplicationsontheAP• Datamining
– Frequentitemset mining– Sequentialpatternmining
• Machinelearning– Randomforest– Entityresolution– String/treekernel
• Bioinformatics– Motifdiscovery– DNAalignment
• …
11/21/17 InternationalConferenceonSupercomputing2017 15
Challenges:ExactFTMontheAP
The AP supports regular languages
Tree can be represented using context-free-grammar [Ivn07]
11/21/17 InternationalConferenceonSupercomputing2017 16
Challenges:ExactFTMontheAP
The AP supports regular languages
Tree can be represented using context-free-grammar [Ivn07]
The AP can not efficiently implement exact FTM
11/21/17 InternationalConferenceonSupercomputing2017 17
Challenges:ExactFTMontheAP
Exact solutions (e.g., stack implementation on the AP)InefficientDatabase dependent Impractical
The AP supports regular languages
Tree can be represented using context-free-grammar [Ivn07]
The AP can not efficiently implement exact FTM
11/21/17 InternationalConferenceonSupercomputing2017 18
Opportunities:Pruning
• Four-stage pruning strategy– Subset pruning– Intersection pruning– Downward pruning – Connectivity pruning
• Kernel properties– Complementary pruning– Avoiding false negatives
Generatecandidatesofk-subtree
PrunecandidatesontheAP
K<maxK&&K-subtreenotempty
K=K+1
Yes
Writeresults
AP
CPU
FindfinalfrequentsubtreeontheCPU
CPU
Isexactsolutionneeded?
No
Yes
No
11/21/17 InternationalConferenceonSupercomputing2017 19
Background• Frequentitemset mining(FIM)
11/21/17 InternationalConferenceonSupercomputing2017 20
Trans. Items
1 Bread,Milk
2 Bread, Diaper, Beer, Eggs
3 Milk,Diaper, Beer, Coke
4 Bread,Milk, Diaper, Beer,Coke
5 Bread, Milk, Coke,Diaper
sup({Diaper,Milk})=3
Trans. Items
1 <{Bread,Milk},{Coke}>
2 <{Bread, Milk,Diaper}{Beer,Eggs}{Diaper}>
3 <{Milk} {Diaper} {Beer, Coke}>
4 <{Bread,Milk, Diaper}{Beer,Diaper}{Beer,Coke,Eggs}>
5 <{Bread, Milk}{Coke}{Diaper}{Eggs}>
• Sequentialpatternmining(SPM)
Bread Milk
Eggs
Kernel1:SubsetPruning• Maingoal:checksdownwardclosureproperty
*Wang,Ke,etal."AssociationruleminingwiththeMicronAutomataProcessor." ParallelandDistributedProcessingSymposium(IPDPS),IEEE,2015.11/21/17 InternationalConferenceonSupercomputing2017 21
Kernel2:IntersectionPruning• Maingoal:checksifallthesubsetsofacandidatehappensinthesametree
11/21/17 InternationalConferenceonSupercomputing2017 22
Kernel3:DownwardPruning• Maingoal:checksifancestordescendantrelationshipismet
* Wang, Ke, Elaheh Sadredini, and Kevin Skadron. "Sequential pattern mining with the Micron automata processor." Proceedings of the ACM International Conference on Computing Frontiers. ACM, 2016.
11/21/17 InternationalConferenceonSupercomputing2017 23
Kernel4:ConnectivityPruning• Maingoal:checksifsiblingrelationshipismet
11/21/17 InternationalConferenceonSupercomputing2017 24
11/21/17 InternationalConferenceonSupercomputing2017 25
Framework
1. MakingARMandSPMautomataforeachkernel2. Creatingappropriateinputstream3. ConfiguringtheautomataforeachkernelontheAPand
streamingthecorrespondinginputstream4. GettingthepotentialfrequentsubtreesfrotheAPoutputafter
applyingthekernels
PerformanceEvaluation• Platform
– CPU:Intel(R)Core™[email protected],Memory:32GB– GPU:TeslaK80,Memory24GB
• Dataset
• Apples-to-applescomparison
11/21/17 InternationalConferenceonSupercomputing2017 26
GPUImplementation• BFSApproach,because:
– NotbeboundbythefiniteGPUglobalmemory– Exposesmanyready-to-processcandidatesandprovideparallelism
• FTM-GPU– CandidategenerationontheCPU– SubsetpruningontheCPU– EnumerationontheGPU
• Treesinsharedmemory• Candidateinconstantmemory
• Sortingtheinputtrees– Decreasedivergence
11/21/17 InternationalConferenceonSupercomputing2017 27
Algorithmic&ArchitecturalContributions
APkerneloverCPUkernelspeedup:Subset=upto163X
Intersection:upto19XDownward:upto3144XConnectivity:upto2635X
11/21/17 InternationalConferenceonSupercomputing2017 28
Algorithmic&ArchitecturalContributions
APkerneloverCPUkernelspeedup:Subset=upto163X
Intersection:upto19XDownward:upto3144XConnectivity:upto2635X
11/21/17 InternationalConferenceonSupercomputing2017 29
Algorithmic&ArchitecturalContributions
APkerneloverCPUkernelspeedup:Subset=upto163X
Intersection:upto19XDownward:upto3144XConnectivity:upto2635X
11/21/17 InternationalConferenceonSupercomputing2017 30
Redbaroverblackbar:upto1.6XBlackbars:upto215XRedbars:upto353X
PruningEfficiency
Kerneleffectiveness:Subset=80%
Intersection:0.5%Downward:3.5%Connectivity:4.8%
11/21/1731
PruningEfficiency
Kerneleffectiveness:Subset=80%
Intersection:0.5%Downward:3.5%Connectivity:4.8%
11/21/1732
Removingintersectionkernel:APoverPatternMatcher:353Xà 2190X
Accuracy:86%à 83%
FTM-APvsOtherFTMAlgorithmsTrade-offbetweenspeed andaccuracy oftheAPsolutionvstheexistingFTMimplementation
33Dataset:CSLOGS
FTM-APvsOtherFTMAlgorithmsTrade-offbetweenspeed andaccuracy oftheAPsolutionvstheexistingFTMimplementation
34Dataset:CSLOGS
FTM-APvsOtherFTMAlgorithmsTrade-offbetweenspeed andaccuracy oftheAPsolutionvstheexistingFTMimplementation
35Dataset:CSLOGS
FTM-APvsOtherFTMAlgorithmsTrade-offbetweenspeed andaccuracy oftheAPsolutionvstheexistingFTMimplementation
36Dataset:CSLOGS
Speedup
&'(_*++,--./0(,-12./
=upto353X
Memoryusage
'/..(<0./=&'(_*+
=upto5000X
ExactSolution:AP+TreeMinerD
11/21/17 InternationalConferenceonSupercomputing2017 37
ExactSolution:AP+TreeMinerD
11/21/17 InternationalConferenceonSupercomputing2017 38
PerformanceEvaluation:ExactSolution:AP+TreeMinerD
4.14X
5.8X
Intel Xeon CPU, 2.30GHz, 512 GB memory, 2.133GHz Dataset:TREEBANK
Summary• Proposeamulti-stagepruningframeworkontheAP
– ThefirstworktousetheAPasapruningmedia– Novelpruningkernels– Abetterscalabilityandstablebehavior– Proposeanexactsolution
11/21/17 InternationalConferenceonSupercomputing2017 40
Takeaways• Rethinkingthealgorithmwhenhavinganewhardwarearchitecture• AppliestospatialautomatacomputingarchitecturesuchasFPGAs• Thisapproachcanbeadoptedforothercomplexpatternminingproblems• Providessomeinsightforthearchitecturalchanges
References• Ke Wang, Elaheh Sadredini, and Kevin Skadron. "Hierarchical Pattern Mining with the Micron Automata Processor.“ International Journal of Parallel
Programming (IJPP), 2017.• Ke Wang, Elaheh Sadredini, and Kevin Skadron. "Sequential Pattern Mining with the Micron Automata Processor." ACM International Conference on
Computing Frontiers, May 2016• J. Wadden, V. Dang, N. Brunelle, T. Tracy II, D. Guo, E. Sadredini, K. Wang, C. Bo, G. Robins, M. Stan, K. Skadron. "ANMLZoo: A Benchmark Suite for
Exploring Bottlenecks in Automata Processing Engines and Architectures." IEEE International Symposium on Workload Characterization (IISWC), October2016.
• Elaheh Sadredini, Reza Rahimi, Ke Wang, and Kevin Skadron. "Frequent Subtree Mining on the Automata Processor: Opportunities and Challenges." ACMInternational Conference on Supercomputing (ICS), Chicago, June 2017
• Shirish Tatikonda, Srinivasan Parthasarathy, and Tahsin Kurc. ”TRIPS and TIDES: new algorithms for tree mining.” Proceedings of the 15th ACMInternational Conference on Information and Knowledge Management. ACM, 2006.
• Ke Wang, Elaheh Sadredini, and Kevin Skadron. ”Sequential pattern mining with the Micron automata processor.” Proceedings of the ACM InternationalConference on Computing Frontiers. ACM, 2016.
• Mohammed Javeed Zaki. ”Efficiently mining frequent trees in a forest: Algorithms and applications.” IEEE Transactions on Knowledge and DataEngineering 17.8 (2005): 1021-1035.
• Yun Chi, et al. ”Frequent subtree mining an overview.” Fundamenta Informaticae 21: 1001-1038, 2011. Paul Dlugosch, et al. ”An efficient and scalablesemiconductor architecture for parallel automata processing.” IEEE Transactions on Parallel and Distributed Systems, 25.12:3088-3098, 2014.
• Renta Ivncsy, and Istvn Vajk. ”Automata Theory Approach for Solving Frequent Pattern Discovery Problems.” World Academy of Science, Engineering andTechnology, International Journal of Computer, Electrical, Automation, Control and Information Engineering 1(8): 2556-2561, 2007.
• Fedja Hadzic, Henry Tan, and Tharam S. Dillon. Mining of data with complex structures. Vol. 333. Springer-Verlag, 2011.• John L. Hennessy, and David A. Patterson. Computer architecture: a quantitative approach. Elsevier, 2011.• Ke Wang, et al. ”Association rule mining with the Micron Automata Processor.” Parallel and Distributed Processing Symposium (IPDPS), IEEE
International, 2015.• Wang, Ke, Kevin Angstadt, Chunkun Bo, Nathan Brunelle, Elaheh Sadredini, Tommy Tracy II, Jack Wadden, Mircea Stan, and Kevin Skadron. "An overview
of micron's automata processor." In Proceedings of the Eleventh IEEE/ACM/IFIP International Conference on Hardware/Software Codesign and SystemSynthesis, p. 14. ACM, 2016.
11/21/17 InternationalConferenceonSupercomputing2017 41
Thank you J
Questions?
11/21/17 InternationalConferenceonSupercomputing2017 42
BackupSlides
11/21/17 InternationalConferenceonSupercomputing2017 43
B
A
D
A
C B
StringEncoding:ABC– BA– – – D
C
A
B
BD
B
A
D
ADCA
D
StringEncoding:AA – CD– BDA– C– – BD– A– – – – BD
PDA-basedSubtreeMining:AnExampleTree
Embeddedsubtree
11/21/17 InternationalConferenceonSupercomputing2017 44
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
DeterministicPDA
11/21/17 InternationalConferenceonSupercomputing2017 45
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:A A – CD– BDA– C– – BD– A– – – – BD
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 46
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
A
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 47
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 48
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – C D– BDA– C– – BD– A– – – – BD
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 49
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD – BDA– C– – BD– A– – – – BD
D
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 50
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 51
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– B DA– C– – BD– A– – – – BD
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 52
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BD A– C– – BD– A– – – – BD
D
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 53
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA – C– – BD– A– – – – BD
A
D
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 54
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
D
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 55
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C – – BD– A– – – – BD
<C,2>
D
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 56
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
D
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 57
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 58
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – B D– A– – – – BD
<B,4>
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 59
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD – A– – – – BD
D
<B,4>
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 60
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
<B,4>
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 61
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A – – – – BD
<A,5>
<B,4>
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 62
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
<B,4>
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 63
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
<B,1>
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 64
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
C
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 65
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 66
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – B D
B
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 67
𝑞?,? 𝒒𝟏,𝟏 𝒒𝟐,𝟐 𝒒𝟑,𝟑
𝒒𝟗,𝟏 𝒒𝟖,𝟐 𝒒𝟕,𝟑 𝒒𝟔,𝟒
𝒒𝟓,𝟑
𝒒𝟒,𝟐A,*/<A,0>* B,*/<B,1>* C,*/<C,2>*
A,*/<A,5>*
-,C/𝜀-,<C,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
Inputtree:AA – CD– BDA– C– – BD– A– – – – BD
<D,9>
B
<A,0>
*
𝜆\A,*/𝜆\A*-,*/𝜀
𝒒𝟏𝟎,𝟏 D,*/<D,9>*
B,*/<B,4>*
-,A/𝜀-,<A,*>/𝜀
-,B/𝜀-,<B,*>/𝜀
𝜆\B,*/𝜆\B*-,*\<A,0>/𝜀
𝜆\C,*/𝜆\C*-,*\<B,1>/𝜀
𝜆,*\𝜆*-,*\{C,<C,*>}/𝜀
𝜆\B,*/𝜆\B*-,*\<B,1>/𝜀
𝜆\A,*/𝜆\A*-,*\<B,4>/𝜀
𝜆,*\𝜆*-,*\{A,<A,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆,*\𝜆*-,*\{B,<B,*>}/𝜀
𝜆\D,*/𝜆\D*-,*\<A,0>/𝜀
𝜆,*/*-,*/𝜀
-,<A,0>/ 𝜀-,<B,1>/ 𝜀
-,<B,1>/ 𝜀
-,<B,4>/ 𝜀
-,<A,0>/ 𝜀
11/21/17 InternationalConferenceonSupercomputing2017 68