Dual-Pivot Quicksort and Beyond: Analysis of Multiway Partitioning and Its Practical Potential

Dual-Pivot Quicksort and BeyondAnalysis of Multiway Partitioning and Its Practical Potential

Sebastian Wild

Vortrag zur wissenschaftlichen Aussprache

im Rahmen des Promotionsverfahrens

8. Juli 2016

Sebastian Wild Dual-Pivot Quicksort and Beyond 2016-07-08 1 / 17

Dual-P??? Q?????? and BeyondAnalysis of Multi??? P??????? and Its Practical Potential

Sebastian Wild



8. Juli 2016


Dual-Pivot Quicksort and BeyondAnalysis of Multiway Partitioning and Its Practical Potential

Sebastian Wild



8. Juli 2016


What is sorting?

What do you think when you hear sorting?


What is sorting?



What is sorting?



What is sorting?



What is sorting?



What is sorting?

Two meanings of sorting

1Separate different sorts

of things

2Put items into order

(alphabetical, numerical, . . . )


What is sorting?



of things




What is sorting?



of things




What is sorting?



of things




Model of Computation

How do computers sort?

1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98

Input given as array of n items

index-based access (5th item A[5] is 10)

we can read and write cells

Only info about data: pairwise comparisons

Is A[5] < A[8]? Yes, 10 < 72

only relative ranking of elements important

may assume, e.g., numbers 1, . . . , n




1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98





Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98

A[5]

10


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98

A[5]

10


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98

A[5]

10

A[8]

72


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98

A[5]

10

A[8]

72


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72






1 2 3 4 5 6 7 8 9

50 2 75 17 10 78 33 72 98

1 2 3 4 5 6 7 8 9

5 1 7 3 2 8 4 6 9


e.g. Excel rows




Is A[5] < A[8]? Yes, 10 < 72




What is Quicksort?

How to make a computer sort?

Selectionsort

Find smallest item and put it in place.

5 1 7 3 2 8 4 6 9


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 915

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

?

1

min

5


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

?

71

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ?

31

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ? ?

21

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ? ? ?

81

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ? ? ? ?

41

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ? ? ? ? ?

61

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ? ? ? ? ? ?

91

min

?


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ? ? ? ? ? ? ?

1

min

5


What is Quicksort?


Selectionsort


5 1 7 3 2 8 4 6 9

? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 95

min

1


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

to do (by same procedure)


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91 75

min

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

?

35

min

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ?

5 3

min


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ?

23

min

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ? ?

2

min

3


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ? ?

82

min

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ? ? ?

42

min

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ? ? ? ?

62

min

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ? ? ? ? ?

92

min

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

5 7 3 2 8 4 6 91

? ? ? ? ? ? ?

2

min

5


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

7 3 5 8 4 6 952

min


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

7 3 5 8 4 6 92


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2 3 7 5 8 4 6 9

? ? ? ? ? ?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

3 4 5 8 7 6 9

? ? ? ? ?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

3 4 5 8 7 6 9

? ? ? ?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Each round puts one item in place.

Remaining list one smaller.

Can we do better?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons



Can we do better?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons



Can we do better?

Yes! Finding the minimum

costs the same as computing

the rank

number of elements

smaller than given element

of any element.


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final place of some item (pivot),

put small items left, others right (partition).


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]

place of some item (pivot),



What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



5 1 7 3 2 8 4 6 9


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



5 1 7 3 2 8 4 6 9


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



5 1 7 3 2 8 4 6 9

k g


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



<

5 1 7 3 2 8 4 6 9

k g

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



<

5 1 7 3 2 8 4 6 9

?

k g


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



<

5 1 7 3 2 8 4 6 9

?

k g

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >

5 1 7 3 2 8 4 6 9

?

k g

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< > >

5 1 7 3 2 8 4 6 9

?

k g

? ?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< > >>

5 1 7 3 2 8 4 6 9

? ? ?

k g

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< > >><

5 1 7 3 2 8 4 6 9

? ? ??

k g

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< >

? ? ??

k g

?

5 1 3 2 8 6 94 7


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< ><

? ? ??

g

?

5 1 4 3 2 8 7 6 9

k g

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< >< <

? ? ??

g

?

5 1 4 3 2 8 7 6 9

?

k g

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< >< < >

? ? ???

5 1 4 3 2 8 7 6 9

? ?

gk

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< >< < >

? ? ???

5 1 4 3 2 8 7 6 9

? ? ?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< >< ><

? ? ???? ? ?

1 4 3 8 7 6 92 5


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< >< ><

? ? ???? ? ?

2 1 4 3 8 7 6 95

reached final place!


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< >>< >< ><

? ? ???? ? ?

2 1 4 3 8 7 6 95

to do to do


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

2 1 4 3 8 7 6 95

to do


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

2 1 4 3 8 7 6 95

to do


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< > >

? ? ???? ? ?

2 1 4 3 8 7 6 95

to do

? ? ?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< > >

? ? ???? ? ?

to do

? ? ?

1 2 4 3 5 8 7 6 9

to do


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

to do

? ? ?

1 2 4 3 5 8 7 6 9

to do


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

to do

? ? ?

1 2 3 4 5 8 7 6 9

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

to do

? ? ?

1 2 3 4 5 8 7 6 9

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< < >

? ? ???? ? ?

? ? ?

1 2 3 4 5 8 7 6 9

?

? ? ?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



< < >

? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9

to do


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9

?


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9

?

16 Comparisons


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9

?

16 Comparisons

Much less comparisons in Quicksort

. . . unless pivots are max/min!

But average/expected behavior is good.

Quicksort is method of choice in practice.


What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9

?

16 Comparisons



But average/expected behavior is good.



What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9

?

16 Comparisons



But average/expected

(actually: almost always good)

behavior is good.



What is Quicksort?


Selectionsort


? ? ? ? ? ? ? ?

1

? ? ? ? ? ? ?

2

? ? ? ? ? ?

? ? ? ? ?

? ? ? ?

3 4 5 6 7 8 9

? ? ?

? ?

?

36 Comparisons

Quicksort

Find final

kth smallest A[k]



? ? ???? ? ?

? ? ?

?

? ? ?

1 2 3 4 5 6 7 8 9

?

16 Comparisons



But average/expected

(actually: almost always good)

behavior is good.



Variants of Quicksort

It is tempting to try to develop ways to improve quicksort:

a faster sorting algorithm is computer science’s “better mousetrap,”

and quicksort is a venerable method that seems to invite tinkering.

R. Sedgewick and K. Wayne, Algorithms 4th ed.







Many variants tried

Two important tuning options:

1 Choose pivots wisely (pivot sampling)

for example: middle element of three (median-of-three)

pivot never min or max!

2 Split into s > 2 subproblems at once (multiway partitioning)

for example: 2 pivots 3 classes: small, medium and large

subproblems are smaller,

but partitioning is more expensive.







Many variants tried















Many variants tried . . . many not helpful































































































































































conventionalwisdom
















conventionalwisdom

till 2009!


Java 7’s Dual-Pivot Quicksort

What happened in 2009? Java

used e.g. for Android apps

7 sorting method was released!

till 2009, Java used classic Quicksort

Java 7 uses new dual-pivot Quicksort

started as hobby project

by Vladimir Yaroslavskiy

20 % faster than

highly-tuned Java 6!










20 % faster than











20 % faster than







till 2009, Java used classic

as in example before

+ median-of-three

+ Insertionsort

Quicksort




20 % faster than









+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn

Java 6 sorting method

(121ms to sort 106 items)

Normalized Java runtimes (in ms).

Average and standard deviation of

1000 random permutations per size.








+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn













+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn













+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn















+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn








No theoretical explanation for running time known in 2009!








+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn









Only lucky experiments?








+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn










Why did no-one come up with this earlier?








+ median-of-three

+ Insertionsort

Quicksort




20 % faster than


0 0.5 1 1.5 2

·106

7

8

9

n

tim

e

10−6·n

lnn










Why did no-one come up with this earlier?

That was the starting point for my work!


Overview of my work

My Thesis

Multiway partitioning has genuine

potential to speed up Quicksort.


Overview of my work

My Thesis



Mathematical Analysisprove eternal truths

not experiments


not experiments


Overview of my work

My Thesis




not experiments

What is multiway partitioning?generalize existing Quicksort variants

generic s-way one-pass partitioning

with pivot sampling



with pivot sampling


Overview of my work

My Thesis




not experiments



with pivot sampling

Unified Analysis of Quicksortcosts on average for large inputs

parametric (algorithm and cost measure)




Overview of my work

My Thesis




not experiments



with pivot sampling



Cost Measures

generalize models of costcomparisonswrite accesses / swapscache misses / scanned elementsbranch mispredictions

element-wise charging schemes

Cost Measures




Overview of my work

My Thesis




not experiments



with pivot sampling



Cost Measures



Discussion of Results

influence ofnumber of pivotspivot samplingorganization of indices




Overview of my work

My Thesis




not experiments



with pivot sampling



Cost Measures






Entropy

What is entropy?

1 Thermodynamics: unusable part of thermal energy

2 Mathematics: information content of a random variable

Definition (Entropy)

Let X be a random variable with P[X = i] = pi for i ∈ {1, . . . ,u}.

Its entropy is defined as H(X) =∑u

i=1 pi log2(1/pi).

probability that X = i

. . . all clear?

Let’s see an application.

1 23

4 56

7 8 9

1011 12

Urn with colored balls.

Alice draws ball X uniformly at random from urn.

Bob wants to guess its color C by Yes/No-Questions.

Theorem (Shannon’s Source Coding Theorem)

On average, Bob can’t do with less than H(C) questions.

We say: We must acquire H(C) bits of information to know C.


Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?

1 Thermodynamics: unusable part of thermal energy (Physicists’ party? Next door, please.)





i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12








Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12



Bob wants to guess its color C

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

by Yes/No-Questions.





Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12




P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25






Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12




P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25




= 4 · 14

log2(4) = 2



Entropy

What is entropy?






i=1 pi log2(1/pi).


. . . all clear?


1 23

4 56

7 8 9

1011 12




P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25

P[C = ] = 3/12 = 0.25




= 4 · 14

log2(4) = 2



Information Theory & Sorting

Where is the connection to sorting?

1 23

4 56

7 8 9

1011 12

Now, Alice draws one of the n!

n(n− 1) · · · 2 · 1

possible orderings of n items.

Bob guesses the order by asking comparisons.

Theorem (Information-Theoretic Lower Bound for Sorting)

We cannot sort with less than log2(n!) = n log2(n)±O(n) comparisons.

Assumptions:

(1) All orderings equally likely.

(2) Only comparisons allowed.asymptotic error term for n → ∞

H(C)

To achieve this bound, every comparison must yield 1 bit of information!




1 23

4 56

7 8 9

1011 12


n(n− 1) · · · 2 · 1

possible orderings of n items.




Assumptions:



H(C)





1 23

4 56

7 8 9

1011 12


n(n− 1) · · · 2 · 1

possible orderings

=̂ colors

of n items.




Assumptions:



H(C)





1 23

4 56

7 8 9

1011 12


n(n− 1) · · · 2 · 1

possible orderings

=̂ colors

of n items.




Assumptions:



H(C)





1 23

4 56

7 8 9

1011 12


n(n− 1) · · · 2 · 1

possible orderings

=̂ colors

of n items.




Assumptions:



H(C)





1 23

4 56

7 8 9

1011 12


n(n− 1) · · · 2 · 1

possible orderings

=̂ colors

of n items.




Assumptions:



H(C)





1 23

4 56

7 8 9

1011 12


n(n− 1) · · · 2 · 1

possible orderings

=̂ colors

of n items.




Assumptions:



H(C)


We can reverse this idea to analyze Quicksort.


Comparisons in Quicksort

How many comparisons does Quicksort use on average?

Entropy-based analysis:

1 Compute the average information yield: x bitcomparison . (0 < x 6 1)

What is x? . . . stay tuned.

2 Need to learn log2(n!) bits Uselog2(n!)

x=

1

xn log2(n)±O(n) comparisons.








x=

1









x=

1









x=

1









x=

1









x=

1








2 Need to learn log2(n!) bits

not mathematically rigorous . . .

but correct result!

rigorous proof → dissertation

Uselog2(n!)

x=

1



Partitioning Entropy

What is x?

Classic Quicksort:

Always compare with the pivot P

< < < < > > > >

2 1 4 3 5 8 7 6 9

Outcome C is either small or large .

Information yield is H(C) =

−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]

prob. of item to be < P

log2(P[ < ]) − P[ > ] log2(P[ > ])



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])

Note: pivot value P is random.

What is its distribution?



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])



Excursion: Input Models

Typical: Random Permutations

items: 1, . . . , n

all orderings equally likely

We use the equivalent

Uniform Model

n numbers drawn i. i.d. Uniform(0, 1)

i. i.d. random relative ranking

P[ < ] = P, P[ > ] = 1− P



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P

same average sorting costs



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P




What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P


independent & identically distributed



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P





What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ]


log2(P[ < ]) − P[ > ] log2(P[ > ])





items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P

0 1P0 1

P[ < ] P[ > ]



Answer: PD= Uniform(0, 1)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P

0 1P0 1

P[ < ] P[ > ]



E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P

0 1P0 1

P[ < ] P[ > ]



E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P

0 1P0 1

P[ < ] P[ > ]



E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



items: 1, . . . , n



Uniform Model



P[ < ] = P, P[ > ] = 1− P

0 1P0 1

P[ < ] P[ > ]



E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348

s-way Quicksort with pivot sampling:

s− 1 pivots: P1, . . . , Ps−1

s classes c1, . . . , cs

Pivot Sampling: parameter~t = (t1, . . . , ts) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)

choose pivots from k = 8 sample items

0 10 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 10 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 10 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 10 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 10 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 10 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 10 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1 U2

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1 U2U3

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1 U2U3 U4

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1 U2U3 U4U5

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1 U2U3 U4U5 U6

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1 U2U3 U4U5 U6 U7

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1

U1 U2U3 U4U5 U6 U7U8

0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1


0 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1


Pivot Sampling: parameter~t = (t1, . . . , ts

no sampling

~t = 0

median-of-3

~t = (1,1)

) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



no sampling

~t = 0

median-of-3

~t = (1,1)

) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

D1 D2 D3

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



no sampling

~t = 0

median-of-3

~t = (1,1)

) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

D1 D2 D3

P[C = c3]

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



no sampling

~t = 0

median-of-3

~t = (1,1)

) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

D1 D2 D3

P[C = c3]

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



no sampling

~t = 0

median-of-3

~t = (1,1)

) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

D1 D2 D3

P[C = c3]

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



no sampling

~t = 0

median-of-3

~t = (1,1)

) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

D1 D2 D3

P[C = c3]

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)

E[H(C)] = information yield of classifying 1 element

If we need a comparisons for that, Quicksort uses

a

E[H(C)]· n log2(n) ± O(n)

comparisons on average.



What is x?

Classic Quicksort:


< < < < > > > >

2 1 4 3 5 8 7 6 9



−P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

E

Average over P

[H(C)] = E

[−P log2(P) − (1− P) log2(1− P)

]

= −2

∫1

0

z log2(z)dz

=1

2log2(e) ≈ 0.721348


s− 1 pivots: P1, . . . , Ps−1



no sampling

~t = 0

median-of-3

~t = (1,1)

) ∈ Ns0

Example: s = 3, ~t = (1, 2, 3)


0 1


P1 P20 1

D1 D2 D3

P[C = c3]

~DD= Dirichlet(t1 + 1, . . . , ts + 1)

E[H(C)] =

s∑

r=1

tr + 1

k+ 1

(Hk+1 −H

1

1+1

2+1

3+ · · ·+ 1

tr+1

tr+1

)· log2(e)

E[H(C)] = information yield of classifying 1 element

If we need a comparisons for that, Quicksort uses

a

E[H(C)]· n log2(n) ± O(n)

comparisons on average.


Binary vs. Multiway Partitioning

Now, does multiway partitioning perform better? Let’s try an example.

4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4

Need a = 2 comparisons to

classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

Classic Quicksort

2n ln(n) comparisons

(without sampling)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

Classic Quicksort


(without sampling)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

Classic Quicksort


(without sampling)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

Classic Quicksort


(without sampling)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

vs.

Classic Quicksort


(without sampling)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

vs.

Classic Quicksort


(without sampling)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

vs.

Classic Quicksort


(without sampling)

Aha! Multiway Quicksort is better!




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

vs.

Classic Quicksort


(without sampling)


127

≈ 1.714

n ln(n) comparisons

(median-of-3)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

vs.

Classic Quicksort


(without sampling)


. . . or not; it depends on sampling.

127

≈ 1.714

n ln(n) comparisons

(median-of-3)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

vs.

Classic Quicksort


(without sampling)



127

≈ 1.714

n ln(n) comparisons

(median-of-3)




4-way Quicksort

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


classify 1 item.

2413

≈ 1.846

n ln(n) comparisons

(without sampling)

vs.

Classic Quicksort


(without sampling)



127

≈ 1.714

n ln(n) comparisons

(median-of-3)

How to separate influence of multiway partitioning from that of sampling?


Entropy-Equivalent Sampling

Key Idea: Simulate by binary partitioning.

One round of 4-way partitioning . . .

(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4

. . . corresponds to three binary partitionings.

< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4

But how to choose pivots P1, P2, P3?





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4


P2 is chosen as median of 3!





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!

can simulate 4-way partitioning

exactly by alternating between

median-of-3 and no sampling.

same on average:1 flip a fair coin

2

{median-of-3 for heads

no sampling for tails

random-parameter pivot sampling





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2








(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2








(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2








(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2








(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2








(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2




Theorem (Entropy-Equivalent Sampling)

For every s-way sampling parameter~t there is a single-pivot

random-parameter scheme with the same quality of pivots.

(formal meaning in dissertation)





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2








Example:~t = (1, 2, 3) with

~T =

{(1, 6) with prob. 9

16

(2, 3) otherwise





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2









~T =


16

(2, 3) otherwise

What does that mean?

1 Can simulate any multiway method by binary partitioning.

2 In terms of comparisons, this can only get better!





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2









~T =


16

(2, 3) otherwise








(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2









~T =


16

(2, 3) otherwise




(with 1 comparison tree)





(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2









~T =


16

(2, 3) otherwise




So, is that the end of multiway Quicksort?






(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2









~T =


16

(2, 3) otherwise





No! . . . stay tuned






(without sampling)

<

< >

>

< >

<P2

<P1

c1 c2

<P3

c3 c4


< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4



P1 and P3 uniform!





2









~T =


16

(2, 3) otherwise





No! . . . stay tuned



The Memory Wall

Comparisons =̂ CPU effort, but we also need to load items from memory.


www.cs.virginia.edu/stream/by_date/Balance.html

The Memory Wall




The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

CPU speed (MFLOPS)

STREAM benchmark data with linear regressionswww.cs.virginia.edu/stream/by_date/Balance.html



The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

CPU speed (MFLOPS)




The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

CPU speed (MFLOPS)


Averaged annual growth rates:

Processor speed: 46%



The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

CPU speed (MFLOPS)

RAM bandwidth (MWords/s)






The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

CPU speed (MFLOPS)





Memory Bandwidth: 37%



The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

CPU speed (MFLOPS)








The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

growing gap

CPU speed (MFLOPS)








The Memory Wall


1991 1995 2000 2005 2010 2015

100

102

104

106

growing gap

CPU speed (MFLOPS)


“imbalance”: speedbandwidth







The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%

Classic Quicksort (Bentley & McIlroy) 1993

1 mem. transfer ≈ 7 operations

optimized for CPU



The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



optimized for CPU



The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



optimized for CPU



The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



optimized for CPU

Dual-Pivot Quicksort, 2009 – today


optimize for CPU and bandwidth?



The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



optimized for CPU






The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



optimized for CPU






The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



optimized for CPU




Why did no-one use dual-pivot Quicksort earlier?

Because it was not faster!



The Memory Wall


1991 1995 2000 2005 2010 2015

100

101

102

CPU speed (MFLOPS)







Imbalance: 5%



optimized for CPU




Why did no-one use dual-pivot Quicksort earlier?

Because it was not faster!



Scanned Elements

How to analyze bandwidth consumption?

Need good cost measure!

Proposal: count scanned elements

Access to array only through iterators

(no random access!)5

1

1

2

7

3

4

4

2

5

8

6

Iterators can

head left or right (one-directional!)

advance to next position

read and write current cell

Cost: number of advances

≈ cache misses, but hardware independent

Costs of Partitioning

Classic Quicksort

k g

n scanned elements

(array scanned once)

4-way Quicksort

k2 g2

k g

1.5n scanned elements

(on average without sampling)


Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g




Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g



4-way simulated by 2-way

< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4

2n scanned elements


Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g




< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4

k g

2n scanned elements


Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g




< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4

k g

k g

2n scanned elements


Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g




< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4

k g

k g k g

2n scanned elements


Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g




< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4

k g

k g k g

2n scanned elements


Scanned Elements






1

1

2

7

3

4

4

2

5

8

6

Iterators can







Classic Quicksort

k g

n scanned elements


4

details on how this

works in one pass

→ in dissertation

-way Quicksort

k2 g2

k g




< >

<P2

c1,2 c3,4

< >

<P1

c1 c2

< >

<P3

c3 c4

k g

k g k g

2n scanned elements

Multiway uses much less scanned elements,

irrespective of pivot sampling.


Conclusion

You have seen:

1

My Thesis



. . . namely: it reduces memory transfers.

2Evidence suggesting high

impact on running time.

3 Idea of mathematical analysis

H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])


Conclusion

You have seen:

1

My Thesis



. . . namely: it reduces memory transfers.




H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])


Conclusion

You have seen:

1

My Thesis


potential to speed up Quicksort.X. . . namely: it reduces memory transfers.




H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])

You have not seen: (in this talk)

Generic multiway partitioning code

k⌈m⌉ k2 k ≡ k1 g ≡ g1 g2 gs−⌊m⌋

s⌈m⌉ . . . s1 ? l1 . . . ls−⌊m⌋

. . . . . .left right

Using two comparison trees

Branch mispredictions

1taken

2taken

3not t.

4not t.

taken

not t.

not t.

not t. not t.

taken

taken

taken

Treatment of equal keys4 2 1 3 3 5 4 4 3 5 2

2 1 2 4 5 4 4 533 3

5 54 442 1 2

5 5

4 2 1 3 3 5 4 4 3 5 2

3

4

5 5

2 1 2

Intriguing interplay

of parameters


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



k⌈m⌉ k2 k ≡ k1 g ≡ g1 g2 gs−⌊m⌋

s⌈m⌉ . . . s1 ? l1 . . . ls−⌊m⌋




1taken

2taken

3not t.

4not t.

taken

not t.

not t.

not t. not t.

taken

taken

taken


2 1 2 4 5 4 4 533 3

5 54 442 1 2

5 5

4 2 1 3 3 5 4 4 3 5 2

3

4

5 5

2 1 2


of parameters


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



k⌈m⌉ k2 k ≡ k1 g ≡ g1 g2 gs−⌊m⌋

s⌈m⌉ . . . s1 ? l1 . . . ls−⌊m⌋




1taken

2taken

3not t.

4not t.

taken

not t.

not t.

not t. not t.

taken

taken

taken


2 1 2 4 5 4 4 533 3

5 54 442 1 2

5 5

4 2 1 3 3 5 4 4 3 5 2

3

4

5 5

2 1 2


of parameters


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



k⌈m⌉ k2 k ≡ k1 g ≡ g1 g2 gs−⌊m⌋

s⌈m⌉ . . . s1 ? l1 . . . ls−⌊m⌋




1taken

2taken

3not t.

4not t.

taken

not t.

not t.

not t. not t.

taken

taken

taken


2 1 2 4 5 4 4 533 3

5 54 442 1 2

5 5

4 2 1 3 3 5 4 4 3 5 2

3

4

5 5

2 1 2


of parameters


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



k⌈m⌉ k2 k ≡ k1 g ≡ g1 g2 gs−⌊m⌋

s⌈m⌉ . . . s1 ? l1 . . . ls−⌊m⌋




1taken

2taken

3not t.

4not t.

taken

not t.

not t.

not t. not t.

taken

taken

taken


2 1 2 4 5 4 4 533 3

5 54 442 1 2

5 5

4 2 1 3 3 5 4 4 3 5 2

3

4

5 5

2 1 2


of parameters


Conclusion

You have seen:

1

My Thesis






H(C) = −P[ < ] log2(P[ < ]) − P[ > ] log2(P[ > ])



k⌈m⌉ k2 k ≡ k1 g ≡ g1 g2 gs−⌊m⌋

s⌈m⌉ . . . s1 ? l1 . . . ls−⌊m⌋




1taken

2taken

3not t.

4not t.

taken

not t.

not t.

not t. not t.

taken

taken

taken


2 1 2 4 5 4 4 533 3

5 54 442 1 2

5 5

4 2 1 3 3 5 4 4 3 5 2

3

4

5 5

2 1 2


of parameters



Image sources

Icons by Freepik and Gregor Cresnar from www.flaticon.com.

Ball Pit: https://www.flickr.com/photos/davidsingleton/2175975083/

Sorting Hat: http://harrypotter.wikia.com/wiki/Sorting_Hat

Highest Towers: WikimediaMouse Trap: bobbleheaddad.comRobert Sedgewick: http://www.cs.princeton.edu/~rs/

Apple: WikimediaOrange: WikimediaCPU Die: WikimediaRAM: Wikimedia


http://www.flaticon.com/authors/freepik

http://www.flaticon.com/authors/gregor-cresnar

www.flaticon.com

https://www.flickr.com/photos/davidsingleton/2175975083/

http://harrypotter.wikia.com/wiki/Sorting_Hat

https://en.wikipedia.org/wiki/List_of_tallest_freestanding_structures#/media/File:Tallest_freestanding_structures_in_the_world.png

http://bobbleheaddad.com/family-life/automatische-wasser-mausefalle-a-read-out-loud-belly-laugh

http://www.cs.princeton.edu/~rs/

https://en.wikipedia.org/wiki/Honeycrisp

https://en.wikipedia.org/wiki/File:Orange-Whole-%26-Split.jpg

https://en.wikipedia.org/wiki/File:GPS_ARM610_die.JPG

https://commons.wikimedia.org/wiki/File:RAM_module_SDRAM_1GiB.jpg

Dual-Pivot Quicksort and Beyond: Analysis of Multiway Partitioning and Its Practical Potential

Science