Download presentation

Presentation is loading. Please wait.

Published byHarry Houghland Modified over 4 years ago

1
Invasion Percolation: Tuning Copyright © Software Carpentry 2010 This work is licensed under the Creative Commons Attribution License See http://software-carpentry.org/license.html for more information. Program Design

2
Invasion PercolationTuning Our program works correctly

3
Program DesignInvasion PercolationTuning Our program works correctly At least, it passes all our tests

4
Program DesignInvasion PercolationTuning Our program works correctly At least, it passes all our tests Next step: figure out if we need to make it faster

5
Program DesignInvasion PercolationTuning Our program works correctly At least, it passes all our tests Next step: figure out if we need to make it faster Spend time on other things if it's fast enough

6
Program DesignInvasion PercolationTuning How do we measure a program's speed?

7
Program DesignInvasion PercolationTuning How do we measure a program's speed? Average of many running times on various grids

8
Program DesignInvasion PercolationTuning How do we measure a program's speed? Average of many running times on various grids How do we predict running time on bigger grids?

9
Program DesignInvasion PercolationTuning How do we measure a program's speed? Average of many running times on various grids How do we predict running time on bigger grids? Use asymptotic analysis

10
Program DesignInvasion PercolationTuning How do we measure a program's speed? Average of many running times on various grids How do we predict running time on bigger grids? Use asymptotic analysis One of the most powerful theoretical tools in computer science

11
Program DesignInvasion PercolationTuning N×N grid has N 2 cells 537261134 856572362 258755659 526493965 468859739 764512685 542585558 575153855 451978651 N N

12
Program DesignInvasion PercolationTuning N×N grid has N 2 cells Must fill at least N to reach boundary 53721134 8565 2362 2587 5659 5264 3965 4688 9739 764512685 542585558 575153855 451978651 N N

13
Program DesignInvasion PercolationTuning N×N grid has N 2 cells Must fill at least N to reach boundary Can fill at most (N-2) 2 +1 = N 2 -4N+5 53721134 8 2 2 9 5 5 4 9 7 5 5 8 5 5 451978651 N N

14
Program DesignInvasion PercolationTuning N×N grid has N 2 cells Must fill at least N to reach boundary Can fill at most (N-2) 2 +1 = N 2 -4N+5 For large N, this is approximately N 2 53721134 8 2 2 9 5 5 4 9 7 5 5 8 5 5 451978651 N N

15
Program DesignInvasion PercolationTuning Program looks at N 2 cells to find the next cell to fill N N

16
Program DesignInvasion PercolationTuning Program looks at N 2 cells to find the next cell to fill Best case: N·N 2 or N 3 steps in total N N

17
Program DesignInvasion PercolationTuning Program looks at N 2 cells to find the next cell to fill Best case: N·N 2 or N 3 steps in total Worst case: N 2 ·N 2 or N 4 steps N N

18
Program DesignInvasion PercolationTuning Program looks at N 2 cells to find the next cell to fill Best case: N·N 2 or N 3 steps in total Worst case: N 2 ·N 2 or N 4 steps Ouch N N

19
Program DesignInvasion PercolationTuning Averaging exponent to N 3.5 doesn't make sense...

20
Program DesignInvasion PercolationTuning Averaging exponent to N 3.5 doesn't make sense......but it will illustrate our problem

21
Program DesignInvasion PercolationTuning Averaging exponent to N 3.5 doesn't make sense......but it will illustrate our problem Grid SizeTime NT 2N11.3 T 3N46.7 T 4N128 T

22
Program DesignInvasion PercolationTuning Grid SizeTimeWhich Is... NT1 sec 2N11.3 T11 sec 3N46.7 T47 sec 4N128 T2 min 10N3162 T52 min 50N883883 T1 day 100N10 7 T115 days Averaging exponent to N 3.5 doesn't make sense......but it will illustrate our problem

23
Program DesignInvasion PercolationTuning Idea #1: Why are we looking at all the cells? N N

24
Program DesignInvasion PercolationTuning Idea #1: Why are we looking at all the cells? Just look at cells that might be adjacent N N

25
Program DesignInvasion PercolationTuning Idea #1: Why are we looking at all the cells? Just look at cells that might be adjacent Keep track of min and max X and Y and loop over those N N

26
Program DesignInvasion PercolationTuning But on average, the filled region is half the size of the grid N N

27
Program DesignInvasion PercolationTuning But on average, the filled region is half the size of the grid N/2 · N/2 = N 2 /4 N N

28
Program DesignInvasion PercolationTuning But on average, the filled region is half the size of the grid N/2 · N/2 = N 2 /4 115 days 29 days N N

29
Program DesignInvasion PercolationTuning But on average, the filled region is half the size of the grid N/2 · N/2 = N 2 /4 115 days 29 days 148N×148N grid eats that up N N

30
Program DesignInvasion PercolationTuning But on average, the filled region is half the size of the grid N/2 · N/2 = N 2 /4 115 days 29 days 148N×148N grid eats that up Need to attack the exponent N N

31
Program DesignInvasion PercolationTuning If we just added this cell... N N

32
Program DesignInvasion PercolationTuning If we just added this cell......why check these cells again? N N

33
Program DesignInvasion PercolationTuning N N If we just added this cell......why check these cells again? We should already know that these are candidates

34
Program DesignInvasion PercolationTuning N N If we just added this cell......why check these cells again? We should already know that these are candidates This is the only new candidate cell

35
Program DesignInvasion PercolationTuning Big idea: trade memory for time

36
Program DesignInvasion PercolationTuning Big idea: trade memory for time Record redundant information in order to save re-calculation

37
Program DesignInvasion PercolationTuning Big idea: trade memory for time Record redundant information in order to save re-calculation In this case, keep a list of cells on the boundary

38
Program DesignInvasion PercolationTuning Big idea: trade memory for time Record redundant information in order to save re-calculation In this case, keep a list of cells on the boundary Each time a cell is filled, check its neighbors

39
Program DesignInvasion PercolationTuning Big idea: trade memory for time Record redundant information in order to save re-calculation In this case, keep a list of cells on the boundary Each time a cell is filled, check its neighbors If already in the list, do nothing

40
Program DesignInvasion PercolationTuning Big idea: trade memory for time Record redundant information in order to save re-calculation In this case, keep a list of cells on the boundary Each time a cell is filled, check its neighbors If already in the list, do nothing Otherwise, add to list

41
Program DesignInvasion PercolationTuning Big idea: trade memory for time Record redundant information in order to save re-calculation In this case, keep a list of cells on the boundary Each time a cell is filled, check its neighbors If already in the list, do nothing Otherwise, add to list Insert into list so that low-valued cells at front

42
Program DesignInvasion PercolationTuning Big idea: trade memory for time Record redundant information in order to save re-calculation In this case, keep a list of cells on the boundary Each time a cell is filled, check its neighbors If already in the list, do nothing Otherwise, add to list Insert into list so that low-valued cells at front Making it easy to choose next cell to fill

43
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 4688597394 7645126853 5425855582 5751538551 4519786510 edge = [] List of cells on edge is initially empty

44
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645126853 5425855582 5751538551 4519786510 edge = [(1,4,3), (8,3,4), (9,4,5), (9,5,4)] List of cells on edge is initially empty Fill center cell and add its neighbors

45
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645126853 5425855582 5751538551 4519786510 edge = [(1,4,3), (8,3,4), (9,4,5), (9,5,4)] value coordinates List of cells on edge is initially empty Fill center cell and add its neighbors

46
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 26853 5425855582 5751538551 4519786510 edge = [(8,3,4), (9,4,5), (9,5,4)] Take first cell from list and fill it

47
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 26853 5425855582 5751538551 4519786510 edge = [(2,5,3), (5,3,3), (8,4,2), (8,3,4), (9,4,5), (9,5,4)] Take first cell from list and fill it Add its neighbors to the list

48
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 6853 5425855582 5751538551 4519786510 edge = [(5,3,3), (5,5,2), (6,6,3), (8,4,2), (8,3,4), (9,4,5), (9,5,4)] Take, fill, and add again

49
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 6853 5425855582 5751538551 4519786510 edge = [(5,3,3), (5,5,2), (6,6,3), (8,4,2), (8,3,4), (9,4,5), (9,5,4)] Take, fill, and add again Don't re-add cells that are already in the list

50
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 6853 5425855582 5751538551 4519786510 edge = [(5,3,3), (5,5,2), (6,6,3), (8,4,2), (8,3,4), (9,4,5), (9,5,4)] In case of ties, find all cells at the front of the list with the lowest value...

51
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 6853 54258 5582 5751538551 4519786510 edge = [(3,5,1), (5,3,3), (5,5,2), (6,6,3), (8,4,2), (8,3,4), (9,4,5), (9,5,4)] In case of ties, find all cells at the front of the list with the lowest value......and select and fill one of them

52
Program DesignInvasion PercolationTuning How quickly can we insert into a sorted list?

53
Program DesignInvasion PercolationTuning def insert(values, new_val): '''Insert (v, x, y) tuple into list in right place.''' i = 0 while i < len(values): if values[i][0] >= new_val[0]: break values.insert(i, new_val) How quickly can we insert into a sorted list?

54
Program DesignInvasion PercolationTuning def insert(values, new_val): '''Insert (v, x, y) tuple into list in right place.''' i = 0 while i < len(values): if values[i][0] >= new_val[0]: break values.insert(i, new_val) How quickly can we insert into a sorted list? Works even if list is empty...

55
Program DesignInvasion PercolationTuning def insert(values, new_val): '''Insert (v, x, y) tuple into list in right place.''' i = 0 while i < len(values): if values[i][0] >= new_val[0]: break values.insert(i, new_val) How quickly can we insert into a sorted list? Works even if list is empty......or new value greater than all values in list

56
Program DesignInvasion PercolationTuning If list has K values...

57
Program DesignInvasion PercolationTuning If list has K values......insertion takes on average K/2 steps

58
Program DesignInvasion PercolationTuning If list has K values......insertion takes on average K/2 steps Our fractals fill about N 1.5 cells...

59
Program DesignInvasion PercolationTuning If list has K values......insertion takes on average K/2 steps Our fractals fill about N 1.5 cells......and there are as many cells on the boundary as there are in the fractal...

60
Program DesignInvasion PercolationTuning If list has K values......insertion takes on average K/2 steps Our fractals fill about N 1.5 cells......and there are as many cells on the boundary as there are in the fractal......so this takes N 1.5 · N 1.5 = N 3 steps

61
Program DesignInvasion PercolationTuning If list has K values......insertion takes on average K/2 steps Our fractals fill about N 1.5 cells......and there are as many cells on the boundary as there are in the fractal......so this takes N 1.5 · N 1.5 = N 3 steps Not much of an improvement on N 3.5

62
Program DesignInvasion PercolationTuning If list has K values......insertion takes on average K/2 steps Our fractals fill about N 1.5 cells......and there are as many cells on the boundary as there are in the fractal......so this takes N 1.5 · N 1.5 = N 3 steps Not much of an improvement on N 3.5 But there's a much faster way to insert

63
Program DesignInvasion PercolationTuning To look up a name in the phone book...

64
Program DesignInvasion PercolationTuning To look up a name in the phone book......open it in the middle

65
Program DesignInvasion PercolationTuning To look up a name in the phone book......open it in the middle If the name there comes after the one you want, go to the middle of the first half

66
Program DesignInvasion PercolationTuning To look up a name in the phone book......open it in the middle If the name there comes after the one you want, go to the middle of the first half If it comes after the one you want, go to the middle of the second half

67
Program DesignInvasion PercolationTuning To look up a name in the phone book......open it in the middle If the name there comes after the one you want, go to the middle of the first half If it comes after the one you want, go to the middle of the second half Then repeat this procedure in that half

68
Program DesignInvasion PercolationTuning To look up a name in the phone book......open it in the middle If the name there comes after the one you want, go to the middle of the first half If it comes after the one you want, go to the middle of the second half Then repeat this procedure in that half 1124456778889

69
Program DesignInvasion PercolationTuning How fast is this? 1124456778889

70
Program DesignInvasion PercolationTuning How fast is this? One probe can find one value 1124456778889

71
Program DesignInvasion PercolationTuning How fast is this? One probe can find one value Two probes can find one value among two (2 1 ) 1124456778889

72
Program DesignInvasion PercolationTuning How fast is this? One probe can find one value Two probes can find one value among two (2 1 ) Three probes can find one value among four (2 2 ) 1124456778889

73
Program DesignInvasion PercolationTuning How fast is this? One probe can find one value Two probes can find one value among two (2 1 ) Three probes can find one value among four (2 2 ) Four probes: one among eight (2 3 ) 1124456778889

74
Program DesignInvasion PercolationTuning How fast is this? One probe can find one value Two probes can find one value among two (2 1 ) Three probes can find one value among four (2 2 ) Four probes: one among eight (2 3 ) K probes: one among 2 K 1124456778889

75
Program DesignInvasion PercolationTuning How fast is this? One probe can find one value Two probes can find one value among two (2 1 ) Three probes can find one value among four (2 2 ) Four probes: one among eight (2 3 ) K probes: one among 2 K log 2 (N) probes: one among N values 1124456778889

76
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 6853 54258 5582 5751538551 4519786510 Running time is N 1.5 · log 2 (N 1.5 ) 1124456778889

77
Program DesignInvasion PercolationTuning 012345678 5372611348 8565723627 2587556596 5264939655 468897394 7645 6853 54258 5582 5751538551 4519786510 Running time is N 1.5 · log 2 (N 1.5 ) Or N 1.5 log 2 (N) if we get rid of constants 1124456778889

78
Program DesignInvasion PercolationTuning That changes things quite a bit Grid SizeOld TimeWhich Was...New TimeWhich Is... NT1 sec 2N11.3 T11 sec2.8 T3 sec 3N46.7 T47 sec8.2 T8 sec 4N128 T2 minutes16 T16 sec 10N3162 T52 minutes105 T2 min 50N883883 T1 day1995 T33 min 100N10 7 T115 days6644 T2 hours

79
Program DesignInvasion PercolationTuning That changes things quite a bit Grid SizeOld TimeWhich Was...New TimeWhich Is... NT1 sec 2N11.3 T11 sec2.8 T3 sec 3N46.7 T47 sec8.2 T8 sec 4N128 T2 minutes16 T16 sec 10N3162 T52 minutes105 T2 min 50N883883 T1 day1995 T33 min 100N10 7 T115 days6644 T2 hours

80
Program DesignInvasion PercolationTuning That changes things quite a bit And the gain gets bigger as N increases Grid SizeOld TimeWhich Was...New TimeWhich Is... NT1 sec 2N11.3 T11 sec2.8 T3 sec 3N46.7 T47 sec8.2 T8 sec 4N128 T2 minutes16 T16 sec 10N3162 T52 minutes105 T2 min 50N883883 T1 day1995 T33 min 100N10 7 T115 days6644 T2 hours

81
Program DesignInvasion PercolationTuning "Divide and conquer" technique used to insert values into list is called binary search

82
Program DesignInvasion PercolationTuning "Divide and conquer" technique used to insert values into list is called binary search Only works if list values are sorted

83
Program DesignInvasion PercolationTuning "Divide and conquer" technique used to insert values into list is called binary search Only works if list values are sorted But it keeps the list values sorted

84
Program DesignInvasion PercolationTuning "Divide and conquer" technique used to insert values into list is called binary search Only works if list values are sorted But it keeps the list values sorted Python implementation is in bisect library

85
Program DesignInvasion PercolationTuning We can do even better

86
Program DesignInvasion PercolationTuning Generating the grid takes N 2 steps 537261134 856572362 258755659 526493965 468859739 764512685 542585558 575153855 451978651 We can do even better

87
Program DesignInvasion PercolationTuning Generating the grid takes N 2 steps If we fill these cells... 537261134 856572362 258755659 526493965 46889739 7645 685 54258 58 57 855 45 978651 We can do even better

88
Program DesignInvasion PercolationTuning Generating the grid takes N 2 steps If we fill these cells......we only ever look at these cells... 537261134 856572362 258755659 526493965 46889739 7645 685 54258 58 57 855 45 978651 We can do even better

89
Program DesignInvasion PercolationTuning Generating the grid takes N 2 steps If we fill these cells......we only ever look at these cells......so why bother generating values for these ones? 9 89 5 6 258 5 7 8 5 9786 We can do even better

90
Program DesignInvasion PercolationTuning Store grid as a dictionary

91
Program DesignInvasion PercolationTuning Store grid as a dictionary Keys are (x,y) coordinates of cells

92
Program DesignInvasion PercolationTuning Store grid as a dictionary Keys are (x,y) coordinates of cells Values are current cell values

93
Program DesignInvasion PercolationTuning Store grid as a dictionary Keys are (x,y) coordinates of cells Values are current cell values Instead of grid[x][y], use get_value(grid, x, y, Z)

94
Program DesignInvasion PercolationTuning Store grid as a dictionary Keys are (x,y) coordinates of cells Values are current cell values Instead of grid[x][y], use get_value(grid, x, y, Z) def get_value(grid, x, y, Z): '''Get value of grid cell, creating if necessary.''' if (x, y) not in grid: grid[(x, y)] = random.randint(1, Z) return grid[(x, y)]

95
Program DesignInvasion PercolationTuning Store grid as a dictionary Keys are (x,y) coordinates of cells Values are current cell values Instead of grid[x][y], use get_value(grid, x, y, Z) def get_value(grid, x, y, Z): '''Get value of grid cell, creating if necessary.''' if (x, y) not in grid: grid[(x, y)] = random.randint(1, Z) return grid[(x, y)] And of course use set_value(grid, x, y, V) as well

96
Program DesignInvasion PercolationTuning Another common optimization

97
Program DesignInvasion PercolationTuning Another common optimization Lazy evaluation

98
Program DesignInvasion PercolationTuning Another common optimization Lazy evaluation Don't compute values until they're actually needed

99
Program DesignInvasion PercolationTuning Another common optimization Lazy evaluation Don't compute values until they're actually needed Again, makes program more complicated...

100
Program DesignInvasion PercolationTuning Another common optimization Lazy evaluation Don't compute values until they're actually needed Again, makes program more complicated......but also faster

101
Program DesignInvasion PercolationTuning Another common optimization Lazy evaluation Don't compute values until they're actually needed Again, makes program more complicated......but also faster Trading human time for machine performance

102
Program DesignInvasion PercolationTuning How much work does this save us?

103
Program DesignInvasion PercolationTuning How much work does this save us? Old cost of creating grid: N 2

104
Program DesignInvasion PercolationTuning How much work does this save us? Old cost of creating grid: N 2 New cost: roughly N 1.5

105
Program DesignInvasion PercolationTuning How much work does this save us? Old cost of creating grid: N 2 New cost: roughly N 1.5 Difference is not N 0.5

106
Program DesignInvasion PercolationTuning How much work does this save us? Old cost of creating grid: N 2 New cost: roughly N 1.5 Difference is not N 0.5 As N gets large, N 2 -N 1.5 N 2

107
Program DesignInvasion PercolationTuning How much work does this save us? Old cost of creating grid: N 2 New cost: roughly N 1.5 Difference is not N 0.5 As N gets large, N 2 -N 1.5 N 2 I.e., without this change, total runtime would be N 2 + N 1.5 log 2 (N) N 2

108
Program DesignInvasion PercolationTuning Moral of the story:

109
Program DesignInvasion PercolationTuning Moral of the story: Biggest performance gains always come from changing algorithms and data structures

110
Program DesignInvasion PercolationTuning The story's other moral:

111
Program DesignInvasion PercolationTuning The story's other moral: Write and test a simple version first, then improve it piece by piece (re-using the tests to check your work)

112
Copyright © Software Carpentry 2010 This work is licensed under the Creative Commons Attribution License See http://software-carpentry.org/license.html for more information. June 2010 created by Greg Wilson

Similar presentations

OK

Aims: To learn about some simple sorting algorithms. To develop understanding of the importance of efficient algorithms. Objectives: All:Understand how.

Aims: To learn about some simple sorting algorithms. To develop understanding of the importance of efficient algorithms. Objectives: All:Understand how.

© 2018 SlidePlayer.com Inc.

All rights reserved.

To make this website work, we log user data and share it with processors. To use this website, you must agree to our Privacy Policy, including cookie policy.

Ads by Google

Ppt on coal petroleum and natural gas Ppt on automobile related topics in economics Ppt on sigma and pi bonds Ppt on chromosomes and genes video Ppt on education for sustainable development Ppt on differential used in automobiles Free ppt on articles in english grammar Ppt on case study format Ppt on instrument landing system ils Ppt on projectile motion