02 analysis

Analysis of Algorithms

‣ estimating running time
‣ mathematical analysis
‣ order-of-growth hypotheses
‣ input models
‣ measuring space
Reference:
Algorithms in Java, Chapter 2
Intro to Programming in Java, Section 4.1 Except as otherwise noted, the content of this presentation
http://www.cs.princeton.edu/algs4 is licensed under the Creative Commons Attribution 2.5 License.

Algorithms in Java, 4th Edition · Robert Sedgewick and Kevin Wayne · Copyright © 2008 · May 2, 2008 10:35:54 AM

Running time

“ As soon as an Analytic Engine exists, it will necessarily guide the
future
course of the science. Whenever any result is sought by its aid,
the question

how many
times do you
have to turn
the crank?

Charles Babbage (1864) Analytic Engine

2

Reasons to analyze algorithms

Predict performance.

this course (COS 226)
Compare algorithms.

Provide guarantees.

Understand theoretical basis. theory of algorithms (COS 423)

client gets poor performance because programmer
did not understand performance characteristics

Primary practical reason: avoid performance bugs.
3

Some algorithmic successes

Discrete Fourier transform.
• Break down waveform of N samples into periodic components.
• Applications: DVD, JPEG, MRI, astrophysics, ….
• Brute force: N2 steps.
Freidrich Gauss
• FFT algorithm: N log N steps, enables new technology. 1805

time
quadratic
64T

32T

16T
linearithmic
8T
linear

size 1K 2K 4K 8K

Linear, linearithmic, and quadratic
4

Some algorithmic successes

N-body Simulation.
• Simulate gravitational interactions among N bodies.
• Brute force: N2 steps.
• Barnes-Hut: N log N steps, enables new research.
Andrew Appel
PU '81

time
quadratic
64T

32T

16T
linearithmic
8T
linear

size 1K 2K 4K 8K

Linear, linearithmic, and quadratic
5

‣ input models
‣ measuring space

6

Scientific analysis of algorithms

A framework for predicting performance and comparing algorithms.

Scientific method.
• Observe some feature of the universe.
• Hypothesize a model that is consistent with observation.
• Predict events using the hypothesis.
• Verify the predictions by making further observations.
• Validate by repeating until the hypothesis and observations agree.

Principles.
• Experiments must be reproducible.
• Hypotheses must be falsifiable.

Universe = computer itself.
7

Experimental algorithmics

Every time you run a program you are doing an experiment!

Why is
my program so slow ?

First step. Debug your program!
Second step. Choose input model for experiments.
Third step. Run and time the program for problems of increasing size.

8

Example: 3-sum

3-sum. Given N integers, find all triples that sum to exactly zero.
Application. Deeply related to problems in computational geometry.

% more 8ints.txt
30 -30 -20 -10 40 0 10 5

% java ThreeSum < 8ints.txt
4
30 -30 0
30 -20 -10
-30 -10 40
-10 0 10

9

3-sum: brute-force algorithm

public class ThreeSum
{
public static int count(long[] a)
{
int N = a.length;
int cnt = 0;
for (int i = 0; i < N; i++)
for (int j = i+1; j < N; j++)
for (int k = j+1; k < N; k++) check each triple

if (a[i] + a[j] + a[k] == 0)
cnt++;
return cnt;
}

public static void main(String[] args)
{
int[] a = StdArrayIO.readLong1D();
StdOut.println(count(a));
}
}

10

Measuring the running time

Q. How to time a program?
A. Manual.

11

Measuring the running time

Q. How to time a program?
A. Automatic.

Stopwatch stopwatch = new Stopwatch();

ThreeSum.count(a);

double time = stopwatch.elapsedTime();
StdOut.println("Running time: " + time + " seconds");

client code

public class Stopwatch
{
private final long start = System.currentTimeMillis();

public double elapsedTime()
{
long now = System.currentTimeMillis();
return (now - start) / 1000.0;
}
}

implementation
12

3-sum: initial observations

Data analysis. Observe and plot running time as a function of input size N.

N time (seconds) †

1024 0.26

2048 2.16

4096 17.18

8192 137.76

† Running Linux on Sun-Fire-X4100

13

Empirical analysis

Log-log plot. Plot running time vs. input size N on log-log scale.

slope = 3

power law

Regression. Fit straight line through data points: c N a. slope

Hypothesis. Running time grows cubically with input size: c N 3.
14

Prediction and verification

Hypothesis. 2.5 × 10 –10 × N 3 seconds for input of size N.

Prediction. 17.18 seconds for N = 4,096.

Observations. N time (seconds)

4096 17.18

4096 17.15 agrees
4096 17.17

Prediction. 1100 seconds for N = 16,384.

Observation. N time (seconds)
agrees
16384 1118.86

15

Doubling hypothesis

Q. What is effect on the running time of doubling the size of the input?

N time (seconds) † ratio

512 0.03 -

1024 0.26 7.88

2048 2.16 8.43

4096 17.18 7.96

8192 137.76 7.96

numbers increases running time increases
by a factor of 2 by a factor of 8
lg of ratio is
exponent in power law
(lg 7.96 ≈ 3)

Bottom line. Quick way to formulate a power law hypothesis.
16

Experimental algorithmics

Many obvious factors affect running time:
• Machine.
• Compiler.
• Algorithm.
• Input data.

More factors (not so obvious):
• Caching.
• Garbage collection.
• Just-in-time compilation.
• CPU use by other applications.

Bad news. It is often difficult to get precise measurements.
Good news. Easier than other sciences.

e.g., can run huge number of experiments

17

‣ input models
‣ measuring space

18

Mathematical models for running time

Total running time: sum of cost × frequency for all operations.
• Need to analyze program to determine set of operations.
• Cost depends on machine, compiler.
• Frequency depends on algorithm, input data.

Donald Knuth
1974 Turing Award

In principle, accurate mathematical models are available.

19

Cost of basic operations

operation example nanoseconds †

integer add a + b 2.1

integer multiply a * b 2.4

integer divide a / b 5.4

floating point add a + b 4.6

floating point multiply a * b 4.2

floating point divide a / b 13.5

sine Math.sin(theta) 91.3

arctangent Math.atan2(y, x) 129.0

... ... ...

† Running OS X on Macbook Pro 2.2GHz with 2GB RAM

20

Cost of basic operations

operation example nanoseconds †

variable declaration int a c1

assignment statement a = b c2

integer compare a < b c3

array element access a[i] c4

array length a.length c5

1D array allocation new int[N] c6 N

2D array allocation new int[N][N] c7 N 2

string length s.length() c8

substring extraction s.substring(N/2, N) c9

string concatenation s + t c10 N

Novice mistake. Abusive string concatenation.
21

Example: 1-sum

Q. How many instructions as a function of N?

int count = 0;
for (int i = 0; i < N; i++)
if (a[i] == 0) count++;

operation frequency

variable declaration 2

assignment statement 2

less than comparison N+1

equal to comparison N
between N (no zeros)
and 2N (all zeros)
array access N

increment ≤ 2N

22

Example: 2-sum


int count = 0;
for (int i = 0; i < N; i++)
for (int j = i+1; j < N; j++)
if (a[i] + a[j] == 0) count++;

1
0 + 1 + 2 + . . . + (N − 1) = N (N − 1)
operation frequency 2
N
=
2
variable declaration N+2

assignment statement N+2

less than comparison 1/2 (N + 1) (N + 2)

equal to comparison 1/2 N (N − 1)
tedious to count exactly
array access N (N − 1)

increment ≤ N2

23

Tilde notation

• Estimate running time (or memory) as a function of input size N.
• Ignore lower order terms.
- when N is large, terms are negligible
- when N is small, we don't care

Ex 1. 6 N 3 + 20 N + 16

~ 6N3
Ex 2. 6 N 3 + 100 N 4/3 + 56
~ 6N3
Ex 3. 6 N 3 + 17 N 2 lg N + 7 N
~ 6N3

discard lower-order terms
(e.g., N = 1000 6 trillion vs. 169 million)

f (N)
Technical definition. f(N) ~ g(N) means lim = 1
N→ ∞ g(N)

€
24

Example: 2-sum

Q. How long will it take as a function of N?

int count = 0;
for (int i = 0; i < N; i++)
for (int j = i+1; j < N; j++)
if (a[i] + a[j] == 0) count++; "inner loop"

operation frequency cost total cost

variable declaration ~ N c1 ~ c1 N

assignment statement ~ N c2 ~ c2 N

less than comparison ~ 1/2 N 2
c3 ~ c3 N 2
equal to comparison ~ 1/2 N 2

array access ~ N2 c4 ~ c4 N 2

increment ≤ N2 c5 ≤ c5 N 2
total ~ cN2

25

Example: 3-sum
478 Algorithms and Data Stru

public class ThreeSum
{ the leading term of mathemati
public static int count(int[] a)
{
pressions by using a mathem
int N = a.length; device known as the tilde no
int cnt = 0; 1
We write f (N) to represe
for (int i = 0; i < N; i++)
quantity that, when divided b
for (int j = i+1; j < N; j++) N
inner
for (int k = j+1; k < N; k++) ~N / 2
2
approaches 1 as N grows. W
loop
if (a[i] + a[j] + a[k] == 0) ~N / 6
3 write N g (N)N (Nf − 1)(N −to indicat
= (N) 2)
3 3!
cnt++; g (N) f (N)1approaches 1 as N
∼ N3
6
return cnt; depends on input data With this notation, we can
}
complicated parts of an exp
public static void main(String[] args) that represent small values. F
{
int[] a = StdArrayIO.readInt1D(); ample, the if statement in Thr
int cnt = count(a); is executed N 3/6 times b
StdOut.println(cnt);
}
N (N 1)(N 2)/6 N 3/6 N
}
Remark. Focus on instructions in inner loop; ignore everything else!N/3, which certainly, when div
Anatomy of a program’s statement execution frequencies N 3/6, approaches 1 as N grow
notation is useful when the ter
ter the leading term are relativ
26

signiﬁcant (for example, when N = 1000, this assumption amounts to sayi
2

Mathematical models for running time


In practice,
• Formulas can be complicated.
• Advanced mathematics might be required.
• Exact models best left for experts.

costs (depend on machine, compiler)

TN = c1 A + c2 B + c3 C + c4 D + c5 E
A = variable declarations
B = assignment statements frequencies
C = compare (depend on algorithm, input)
D = array access
E = increment

Bottom line. We use approximate models in this course: TN ~ c N3.
27

‣ input models
‣ measuring space

28

Common order-of-growth hypotheses

To determine order-of-growth:
• Assume a power law TN ~ c N a.
• Estimate exponent a with doubling hypothesis.
• Validate with mathematical analysis.

N time (seconds) †
Ex. ThreeSumDeluxe.java
1,000 0.43
Food for thought. How is it implemented?
2,000 0.53

4,000 1.01

8,000 2.87

16,000 11.00

32,000 44.64

64,000 177.48

observations

Caveat. Can't identify logarithmic factors with doubling hypothesis.

29

482 Algorithms and Data Structures

Linearithmic. We use the term linearithmic to describe programs whose running
time for a problem of size N has order of growth N log N. Again, the base of the
Good news. the smallrelevant. For example, CouponCollector (PROGRAM 1.4.2) is lin-
logarithm is not set of functions
earithmic.N, prototypical example 2, mergesort (see PROGRAM 4.2.6). Several im-
1, log The N, N log N, N is N3, and 2N
portant problems have natural solutions that are quadratic but clever algorithms
suffices to describe order-of-growth of typical algorithms. impor-
that are linearithmic. Such algorithms (including mergesort) are critically
tant in practice because they enable us to address problem sizes far larger than
could be addressed with quadratic solutions. In the next section, we consider a
general design technique for developing
time growth rate
linearithmic algorithms. name T(2N) / T(N)

1024T
Quadratic. A 1 constant 1
exponential

typical program whose
c

tic
cubi

lin mic
dra

512T

h
running time has order of growth N 2

r
rit

ea
log N logarithmic ~1
qua

ea
lin has two nested for loops, used for some
calculation involving all pairs linear ele-
N of N 2
64T ments. The force update double loop in
NBody (PROGRAM 3.4.2) is a linearithmicof
N log N prototype ~2
the programs in this classiﬁcation, as is
the elementaryN2 quadratic
sorting algorithm Inser- 4
8T tion (PROGRAM 4.2.4).
4T
N 3
cubic 8

2T Cubic. Our example for this section,
logarithmic
N
2 exponential T(N)
ThreeSum,
is cubic (its running time has
T
constant order of growth N 3) because it has three
size 1K 2K 4K 8K 1024K
nested for loops, to process all triples of
N elements. The running time of matrix factor for
Orders of growth (log-log plot)
multiplication, as implemented in SEC- doubling hypothesis
TION 1.4 has order of growth M 3 to mul-
tiply two M-by-M matrices, so the basic matrix multiplication algorithm is often
considered to be cubic. However, the size of the input (the number of entries in the 30
matrices) is proportional to N = M 2, so the algorithm is best classiﬁed as N 3/2, not
cubic.


growth
name typical code framework description example
rate

1 constant a = b + c; statement add two numbers

while (N > 1)
log N logarithmic { N = N / 2; ... } divide in half binary search

for (int i = 0; i < N; i++)
N linear { ... } loop find the maximum

divide
N log N linearithmic [see lecture 5] mergesort
and conquer

for (int i = 0; i < N; i++)
N2
quadratic for (int j = 0; j < N; j++) double loop check all pairs
{ ... }

for (int i = 0; i < N; i++)
for (int j = 0; j < N; j++)
N3 cubic for (int k = 0; k < N; k++) triple loop check all triples
{ ... }

exhaustive check all
2N exponential [see lecture 24]
search possibilities

31

Practical implications of order-of-growth

Q. How long to process millions of inputs?
Ex. Population of NYC was "millions" in 1970s; still is.
seconds equivalent

1 1 second
Q. How many inputs can be processed in minutes?
10 10 seconds
Ex. Customers lost patience waiting "minutes" in 1970s;
102 1.7 minutes
they still do.
103 17 minutes

104 2.8 hours

105 1.1 days
For back-of-envelope calculations, assume:
106 1.6 weeks

107 3.8 months
processor instructions
decade
speed per second 108 3.1 years

1970s 1 MHz 106 109 3.1 decades

1980s 10 MHz 107 1010 3.1 centuries

1990s 100 MHz 108 … forever

2000s 1 GHz 109 1017 age of universe

32


problem size solvable in minutes time to process millions of inputs
growth
rate
1970s 1980s 1990s 2000s 1970s 1980s 1990s 2000s

1 any any any any instant instant instant instant

log N any any any any instant instant instant instant

tens of hundreds of
N millions billions minutes seconds second instant
millions millions

hundreds of hundreds of tens of
N log N millions millions hour minutes seconds
thousands millions seconds

tens of
N2 hundreds thousand thousands decades years months weeks
thousands

N3 hundred hundreds thousand thousands never never never millennia

33


effect on a program that
runs for a few seconds
growth
name description
rate
time for 100x size for 100x
more data faster computer

1 constant independent of input size - -

log N logarithmic nearly independent of input size - -

N linear optimal for N inputs a few minutes 100x

N log N linearithmic nearly optimal for N inputs a few minutes 100x

N2 quadratic not practical for large problems several hours 10x

N3 cubic not practical for medium problems several weeks 4-5x

2N exponential useful only for tiny problems forever 1x

34

‣ input models
‣ measuring space

35

Types of analyses

Best case. Running time determined by easiest inputs.
Ex. N-1 compares to insertion sort N elements in ascending order.

Worst case. Running time guarantee for all inputs.
Ex. No more than ½N2 compares to insertion sort any N elements.

Average case. Expected running time for "random" input.
Ex. ~ ¼ N2 compares on average to insertion sort N random elements.

36

Commonly-used notations

notation provides example shorthand for used to

10 N 2
provide
Tilde leading term ~ 10 N 2 10 N 2 + 22 N log N
approximate model
10 N 2 + 2 N +37

N2
asymptotic classify
Big Theta Θ(N 2) 9000 N 2
growth rate algorithms
5 N 2 + 22 N log N + 3N

N2
develop
Big Oh Θ(N 2) and smaller O(N 2) 100 N
upper bounds
22 N log N + 3 N

9000 N 2
develop
Big Omega Θ(N 2) and larger Ω(N 2) N5
lower bounds
N 3 + 22 N log N + 3 N

37

Commonly-used notations

Ex 1. Our brute-force 3-sum algorithm takes Θ(N 3) time.

Ex 2. Conjecture: worst-case running time for any 3-sum algorithm is Ω(N 2).

Ex 3. Insertion sort uses O(N 2) compares to sort any array of N elements;
it uses ~ N compares in best case (already sorted) and ~ ½N 2 compares in the
worst case (reverse sorted).

Ex 4. The worst-case height of a tree created with union find with path
compression is Θ(N).

Ex 5. The height of a tree created with weighted quick union is O(log N).

base of logarithm absorbed by big-Oh

1
loga N = logb N
logb a

38

Predictions and guarantees

Theory of algorithms. Worst-case running time of an algorithm is O(f(N)).

Advantages
• describes guaranteed performance. time/memory

• O-notation absorbs input model.

f(N)
Challenges values represented

•
by O(f(N))
Cannot use to predict performance.
• Cannot use to compare algorithms.

input size

39

Predictions and guarantees

Experimental algorithmics. Given input model,average-case
running time is ~ c f(N).

Advantages. time/memory

• Can use to predict performance.
• Can use to compare algorithms.
c f(N)

values represented
Challenges. by ~ c f(N)

• Need to develop accurate input model.
• May not provide guarantees.

input size

40

‣ input models
‣ measuring space

41

Typical memory requirements for primitive types in Java

Bit. 0 or 1.
Byte. 8 bits.
Megabyte (MB). 210 bytes ~ 1 million bytes.
Gigabyte (GB). 220 bytes ~ 1 billion bytes.

type bytes

boolean 1

byte 1

char 2

int 4

float 4

long 8

double 8

42

Typical memory requirements for arrays in Java

Array overhead. 16 bytes on a typical machine.

type bytes type bytes

char[] 2N + 16 char[][] 2N2 + 20N + 16

int[] 4N + 16 int[][] 4N2 + 20N + 16

double[] 8N + 16 double[][] 8N2 + 20N + 16

one-dimensional arrays two-dimensional arrays

Q. What's the biggest double[] array you can store on your computer?

typical computer in 2008 has about 1GB memory

43

M 3.2.1) object uses 32
ee double instance vari-
Typical memory requirements for objects in Java
ny programs create mil-
the information needed
Object overhead. 8 bytes on a typical machine.
or transparency). A refer-
Reference. 4 bytes on a typical machine.
n a data type contains a
4 bytes for theEach Complex object consumes 24 bytes of memory.
Ex 1.
reference
y needed for the object’s
(Program 3.2.1) public class Complex
32 bytes 8 bytes overhead for object

class Charge {
object
private double re; 8 bytes
overhead
ate double rx; private double im; 8 bytes
ate double ry; rx
... double
ate double q; ry
} values 24 bytes
q

ct (Program 3.2.6) 24 bytes
class Complex
object
overhead
ate double re;
ate double im; re double
im values

Java library) 12 bytes 44
class Color
object
overhead
ate byte r;

...
values
q
}

omplex object (Program 3.2.6) 24 bytes
public class Typical memory requirements for objects in Java
Complex
{ object
overhead
private double re;
private double im; re double
... Object overhead. im bytes on a
8 values typical machine.
}
Reference. 4 bytes on a typical machine.
olor object (Java library) 12 bytes
public class Color
{ Ex 2. A String of length
object
overhead N consumes 2N + 40 bytes.
private byte r;
private byte g; r g b a
private byte b;
private byte a;
... public class String
byte 8 bytes overhead for object
} values
{
private int offset; 4 bytes
ocument object (Program 3.3.4)
16 bytes + string + vector
private int count; 4 bytes
public class Document
{ private int hash;
object 4 bytes
overhead
private String id;private char[] value; 4 bytes for reference
private Vector profile; id
... references (plus 2N + 16 bytes for array)
... profile
} }
2N + 40 bytes
ring object (Java library) 24 bytes + char array
public class String
{ object
overhead
private char[] val;
private int offset; value reference
private int count; offset
private int hash; int
count values
...
} hash
45

Typical object memory requirements

Example 1

Q. How much memory does this program use as a function of N ?

public class RandomWalk {
public static void main(String[] args) {
int N = Integer.parseInt(args[0]);
int[][] count = new int[N][N];
int x = N/2;
int y = N/2;

for (int i = 0; i < N; i++) {
// no new variable declared in loop
...
count[x][y]++;
}
}
}

46

Example 2

Q. How much memory does this code fragment use as a function of N ?

...
int N = Integer.parseInt(args[0]);
for (int i = 0; i < N; i++) {
int[] a = new int[N];
...
}

Remark. Java automatically reclaims memory when it is no longer in use.

47

Out of memory

Q. What if I run out of memory?

% java RandomWalk 10000
Exception in thread "main" java.lang.OutOfMemoryError: Java heap space

% java -Xmx 500m RandomWalk 10000
...

% java RandomWalk 30000
Exception in thread "main" java.lang.OutOfMemoryError: Java heap space

% java -Xmx 4500m RandomWalk 30000
Invalid maximum heap size: -Xmx4500m
The specified size exceeds the maximum representable size.
Could not create the Java virtual machine.

48

Turning the crank: summary

In practice, approximate mathematical models are easily achieved.

Timing may be flawed?
• Limits on experiments insignificant compared to
other sciences.

• Mathematics might be difficult?
• Only a few functions seem to turn up.
• Doubling hypothesis cancels complicated constants.

Actual data might not match input model?
• Need to understand input to effectively process it.
• Approach 1: design for the worst case.
• Approach 2: randomize, depend on probabilistic guarantee.

49

02 analysis

More Related Content

What's hot (20)

Similar to 02 analysis (20)

More from Rick Boardman (11)

Recently uploaded (20)

02 analysis