1. Drools Planner:
Planning optimization
by example
Geoffrey De Smet

2. Demo
Drools Planner
TSP example
2

3. Every organization has planning problems.
3

4. What is a planning problem?
 Complete goals
 With limited resources
 Under constraints
4

5. Hospital bed scheduling
 Assign each
● Patient
 To
● Bed
 Constraints
http://www.flickr.com/photos/markhillary/2227726759/
● Length of stay
● Room requirements
● Room preferences
5

6. Hospital nurse rostering
 Assign each
● Shift
 To
● Nurse
 Constraints
http://www.flickr.com/photos/glenpooh/709704564/
● Employment
contract
● Free time
preferences
● Skills
6

7. School timetabling
 Assign each
● Course
 To
● Room
● Timeslot
http://www.flickr.com/photos/phelyan/2281095105/
 Constraints
● No simultaneous
courses
● Per room
● Per teacher
● Per student
7

8. Airline routing
 Assign each
● Flight
 To
● Airplane
● Crew
http://www.flickr.com/photos/yorickr/3674349657/
 Constraints
● Airplane/crew
depart from where
they arrive
● Minimize mileage
8

9. Bin packing in the cloud
 Assign each
● Process
 To
● Server
 Constraints
● Hardware
http://www.flickr.com/photos/torkildr/3462607995/
requirements
● Minimize server cost
9

10. Bin packing in the cloud
 Assign each
● Process
 To
● Server
 Constraints
● Hardware
http://www.flickr.com/photos/torkildr/3462607995/
requirements
● Minimize server cost
10

11. Which processes do we assign
to this server?
11

12. How did we find that solution?
12

13. First fit by decreasing size
13

14. First fit by decreasing size
14

15. First fit by decreasing size
15

16. First fit by decreasing size
16

17. Another case
17

18. Try FFD again
18

19. Try FFD again
19

20. Try FFD again
20

21. Try FFD again
21

22. FFD failure
22

23. NP complete
23

24. NP complete
 No silver bullet known
● Holy grail of
computer science
● P == NP
● Probably does not
exist
● P != NP
 Root problem of
http://www.flickr.com/photos/annguyenphotography/3267723713/
all planning
problems
24

25. Multiple servers...
25

26. Multiple servers...
26

27. Multiple servers...
27

28. Multiple servers...
28

29. Multiple constraints...
29

30. Multiple constraints...
30

31. Multiple constraints...
31

32. Multiple constraints...
32

33. Organizations rarely optimize
planning problems.
33

34. Reuse optimization algorithms?
34

35.  Open source (ASL 2.0)
 Regular releases
 Reference manual
 Examples
35

36. Given 2 solutions, which one is better?
36

37. Each Solution has 1 Score
37

38. Each Solution has 1 Score
38

39. Better score => better solution
39

40. Better score => better solution
40

41. Best score => best solution
41

42. Best score => best solution
42

43. Demo
Drools Planner
CloudBalance example
43

44. Domain model

45. Server
45

46. Server
public class Server {
private int cpuPower;
private int memory;
private int networkBandwidth;
private int cost;
// getters
}
46

47. Process
47

48. Process
48

49. Process is a planning entity
@PlanningEntity
public class Process {
private int requiredCpuPower;
private int requiredMemory;
private int requiredNetworkBandwidth;
...
// getters, clone, equals, hashcode
}
49

50. Process has a
planning variable
@PlanningEntity
public class Process {
...
private Server server;
@PlanningVariable
@ValueRange(type = FROM_SOLUTION_PROPERTY,
solutionProperty = "serverList")
public Server getServer() {
return server;
}
public void setServer(Server server) {...}
...
}
50

51. CloudBalance
51

52. Solution CloudBalance:
problem facts
public class CloudBalance
implements Solution<HardAndSoftScore> {
private List<Server> serverList;
public List<Server> getServerList() {
return serverList;
}
...
}
52

53. Solution CloudBalance:
planning entities
public class CloudBalance
implements Solution<HardAndSoftScore> {
...
private List<Process> processList;
@PlanningEntityCollectionProperty
public List<Process> getProcessList() {
return processList;
}
...
}
53

54. Solution CloudBalance:
getProblemFacts
public class CloudBalance
implements Solution<HardAndSoftScore> {
...
// Used in score constraints
public Collection<Object> getProblemFacts() {
List<Object> facts = new ArrayList<Object>();
facts.addAll(serverList);
return facts;
}
...
}
54

55. Solution CloudBalance:
score
public class CloudBalance
implements Solution<HardAndSoftScore> {
...
private HardAndSoftScore score;
public HardAndSoftScore getScore() {...}
public void setScore(HardAndSoftScore score) {...}
...
}
55

56. Score constraints

57. Score calculation
with Drools rule engine
 DRL
● Declarative
● Like SQL, regular expressions
 Performance + scalability
● Indexing, ReteOO, ...
● Incremental (delta based) score calculation
● Solution changes => only recalculate part of
score
57

58. Soft constraint:
server cost
rule "serverCost"
when
// there is a server
$s : Server($c : cost)
// there is a processes on that server
exists Process(server == $s)
then
// lower soft score by $c
insertLogical(new IntConstraintOccurrence(
"serverCost", ConstraintType.NEGATIVE_SOFT,
$c,
$s));
end
58

59. Hard constraint:
CPU power
rule "requiredCpuPowerTotal"
when
// there is a server
$s : Server($cpu : cpuPower)
// with too little cpu for its processes
$total : Number(intValue > $cpu) from accumulate(
Process(server == $s,
$requiredCpu : requiredCpuPower),
sum($requiredCpu)
)
then
// lower hard score by ($total.intValue() - $cpu)
end
59

60. Solving it

61. Solver configuration by XML
<solver>
<solutionClass>...CloudBalance</solutionClass>
<planningEntityClass>...Process</>
<scoreDrl>...ScoreRules.drl</scoreDrl>
<scoreDefinition>
<scoreDefinitionType>HARD_AND_SOFT</>
</scoreDefinition>
<!-- optimization algorithms -->
</solver>
61

62. Solving
XmlSolverFactory factory = new XmlSolverFactory(
"...SolverConfig.xml");
Solver solver = factory.buildSolver();
solver.setPlanningProblem(cloudBalance);
solver.solve();
cloudBalance = (CloudBalance)
solver.getBestSolution();
62

63. Optimization algorithms

64. Brute Force
64

65. Brute Force config
<solver>
...
<bruteForce />
</solver>
65

66. Brute force scalability
9 processes = 1.5 seconds
* 100
6 processes = 15 ms
66

67. Brute force scalability
12 processes = 17 minutes
* 680
9 processes = 1.5 seconds
6 processes = 15 ms * 100
67

68. Plan 1200 processes
with brute force?
68

69. First Fit
69

70. First Fit config
<solver>
...
<constructionHeuristic>
<constructionHeuristicType>FIRST_FIT</>
</constructionHeuristic>
</solver>
70

71. First Fit scalability
71

72. First Fit results
CPU: OK
RAM: OK
Network: OK
Cost active servers: shown
72

73. First Fit Decreasing
73

74. First Fit Decreasing config
<solver>
...
<constructionHeuristic>
<constructionHeuristicType>FIRST_FIT_DECREASING</>
</constructionHeuristic>
</solver>
74

75. DifficultyComparator
public class ProcessDifficultyComparator
implements Comparator<Process> {
public int compare(Process a, Process b) {
// Compare on requiredCpuPower * requiredMemory
// * requiredNetworkBandwidth
}
}
@PlanningEntity(difficultyComparatorClass
= ProcessDifficultyComparator.class)
public class Process {
...
}
75

76. First Fit Decreasing scalability
76

77. First Fit Decreasing results
77

78. Local Search
78

79. Local Search comes after
Construction Heuristics
<solver>
...
<constructionHeuristic>
<constructionHeuristicType>FIRST_FIT_DECREASING</>
</constructionHeuristic>
<localSearch>
...
<localSearch>
</solver>
79

80. Local Search needs to be
terminated
<solver>
...
<termination>
<maximumMinutesSpend>20</maximumMinutesSpend>
</termination>
...
</solver>
80

81. Selecting moves
<localSearch>
<selector>
<selector>
<moveFactoryClass>GenericChangeMoveFactory</>
</selector>
<selector>
<moveFactoryClass>GenericSwapMoveFactory</>
</selector>
</selector>
... tabu search, simulated annealing or ...
</localSearch>
81

82. Hill climbing
 Select the best improving move
 Gets stuck in local optima
● Not good
82

83. Hill climbing
<localSearch>
<selector>...</selector>
<forager>
<!-- Untweaked standard value -->
<minimalAcceptedSelection>1000</>
</forager>
</localSearch>
83

84. Tabu search
 Hill climbing ++
 Select the best move
● Recently selected moves are taboo
● Even if the best move is not improving
 Does not get stuck in local optima
84

85. Tabu Search
<localSearch>
<selector>...</selector>
<acceptor>
<!-- Untweaked standard value -->
<propertyTabuSize>7</propertyTabuSize>
</acceptor>
<forager>
<!-- Untweaked standard value -->
<minimalAcceptedSelection>1000</>
</forager>
</localSearch>
85

86. Simulated annealing
 Randomly select moves
● Prefer improving moves
● Especially as time goes by
 Does not get stuck in local optima
86

87. Simulated Annealing
<localSearch>
<selector>...</selector>
<acceptor>
<!-- Tweaked value -->
<simulatedAnnealingStartingTemperature>
0hard/400soft</>
</acceptor>
<forager>
<!-- Untweaked standard value -->
<minimalAcceptedSelection>4</>
</forager>
</localSearch>
87

88. Local Search results
88

89. Cost ($) reduction
 Compared to First Fit
● First Fit Decreasing
● 49 480 $ = 4 %
● Tabu Search
● 160 860 $ = 14
%
● Simulated annealing
● 128 950 $ = 11
%
 Few constraints here
● => low ROI
89

90. Organizations rarely optimize
planning problems.
90

91. http://www.flickr.com/photos/techbirmingham/345897594/
Organizations waste resources.
91

92. Real-time planning

93. Real-time paradox
 First Fit Decreasing 1200 processes = 8 seconds
 Time allowed: 100ms after last change
93

94. Demo
Real-time planning
CloudBalance example
94

95. Demo
Real-time planning
TSP example
95

96. Summary

97. Summary
 Drools Planner optimizes planning
 Adding constraints is easy and scalable
 Switching/combining algorithms is easy
97

98. Try an example!
98

99. Q&A
 Drools Planner homepage:
● http://www.jboss.org/drools/drools-planner
 Reference manual:
● http://www.jboss.org/drools/documentation
 Download this presentation:
● http://www.jboss.org/drools/presentations


 Twitter: @geoffreydesmet
 Google+: Geoffrey De Smet
99