Slide 1

Back to George One More Time Before they invented drawing boards, what did they go back to? If all the world is a stage, where is the audience sitting? If the #2 pencil is the most popular, why is it still #2? If work is so terrific, how come they have to pay you to do it? If you ate pasta and antipasto, would you still be hungry? If you try to fail, and succeed, which have you done? "People who think they know everything are a great annoyance to those of us who do. - Anon

Slide 2

O() Analysis Reasonable vs. Unreasonable Algorithms Using O() Analysis in Design Concurrent Systems Parallelism

Slide 3

Recipe for Determining O() Break algorithm down into known pieces Well learn the Big-Os in this section Identify relationships between pieces Sequential is additive Nested (loop / recursion) is multiplicative Drop constants Keep only dominant factor for each variable

Slide 4

Comparing Data Structures and Methods Data StructureTraverseSearchInsert Unsorted L ListNN1 Sorted L ListNNN Unsorted ArrayNN1 Sorted ArrayNLog NN Binary TreeNN1 BSTNNN F&B BSTNLog NLog N LB

Slide 5

Reasonable vs. Unreasonable Algorithms

Slide 6

Algorithmic Performance Thus Far Some examples thus far: O(1) Insert to front of linked list O(N) Simple/Linear Search O(N Log N)MergeSort O(N 2 )BubbleSort But it could get worse: O(N 5 ), O(N 2000 ), etc.

Slide 7

An O(N 5 ) Example For N = 256 N 5 = 256 5 = 1,100,000,000,000 If we had a computer that could execute a million instructions per second 1,100,000 seconds = 12.7 days to complete But it could get worse

Slide 8

The Power of Exponents A rich king and a wise peasant

Slide 9

The Wise Peasants Pay Day(N)Pieces of Grain 12 24 38 416... 2N2N 63 9,223,000,000,000,000,000 64 18,450,000,000,000,000,000

Slide 10

How Bad is 2 N ? Imagine being able to grow a billion (1,000,000,000) pieces of grain a second It would take 585 years to grow enough grain just for the 64 th day Over a thousand years to fulfill the peasants request!

Slide 11

So the King cut off the peasants head. LB

Slide 12

The Towers of Hanoi A B C Goal: Move stack of rings to another peg Rule 1: May move only 1 ring at a time Rule 2: May never have larger ring on top of smaller ring

Slide 13

The Towers of Hanoi A B C

Slide 14

The Towers of Hanoi A B C

Slide 15

The Towers of Hanoi A B C

Slide 16

The Towers of Hanoi A B C

Slide 17

The Towers of Hanoi A B C

Slide 18

The Towers of Hanoi A B C

Slide 19

The Towers of Hanoi A B C

Slide 20

The Towers of Hanoi A B C

Slide 21

The Towers of Hanoi A B C

Slide 22

The Towers of Hanoi A B C

Slide 23

The Towers of Hanoi A B C

Slide 24

The Towers of Hanoi A B C

Slide 25

The Towers of Hanoi A B C

Slide 26

The Towers of Hanoi A B C

Slide 27

The Towers of Hanoi A B C

Slide 28

The Towers of Hanoi A B C

Slide 29

Towers of Hanoi - Complexity For 1 rings we have 1 operations. For 2 rings we have 3 operations. For 3 rings we have 7 operations. For 4 rings we have 15 operations. In general, the cost is 2 N 1 = O(2 N ) Each time we increment N, we double the amount of work. This grows incredibly fast!

Slide 30

Towers of Hanoi (2 N ) Runtime For N = 64 2 N = 2 64 = 18,450,000,000,000,000,000 If we had a computer that could execute a million instructions per second It would take 584,000 years to complete But it could get worse

Slide 31

The Bounded Tile Problem Match up the patterns in the tiles. Can it be done, yes or no?

Slide 32

The Bounded Tile Problem Matching tiles

Slide 33

Tiling a 5x5 Area 25 available tiles remaining

Slide 34

Tiling a 5x5 Area 24 available tiles remaining

Slide 35

Tiling a 5x5 Area 23 available tiles remaining

Slide 36

Tiling a 5x5 Area 22 available tiles remaining

Slide 37

Tiling a 5x5 Area 2 available tiles remaining

Slide 38

Analysis of the Bounded Tiling Problem Tile a 5 by 5 area (N = 25 tiles) 1st location: 25 choices 2nd location: 24 choices And so on Total number of arrangements: 25 * 24 * 23 * 22 * 21 *.... * 3 * 2 * 1 25! (Factorial) = 15,500,000,000,000,000,000,000,000 Bounded Tiling Problem is O(N!)

Slide 39

Tiling (N!) Runtime For N = 25 25! = 15,500,000,000,000,000,000,000,000 If we could place a million tiles per second It would take 470 billion years to complete Why not a faster computer?

Slide 40

A Faster Computer If we had a computer that could execute a trillion instructions per second (a million times faster than our MIPS computer) 5x5 tiling problem would take 470,000 years 64-ring Tower of Hanoi problem would take 213 days Why not an even faster computer!

Slide 41

The Fastest Computer Possible? What if: Instructions took ZERO time to execute CPU registers could be loaded at the speed of light These algorithms are still unreasonable! The speed of light is only so fast!

Slide 42

Where Does this Leave Us? Clearly algorithms have varying runtimes. Wed like a way to categorize them: Reasonable, so it may be useful Unreasonable, so why bother running

Slide 43

Performance Categories of Algorithms Sub-linear O(Log N) Linear O(N) Nearly linear O(N Log N) Quadratic O(N 2 ) Exponential O(2 N ) O(N!) O(N N ) Polynomial

Slide 44

Reasonable vs. Unreasonable Reasonable algorithms have polynomial factors O (Log N) O (N) O (N K ) where K is a constant Unreasonable algorithms have exponential factors O (2 N ) O (N!) O (N N )

Slide 45

Reasonable vs. Unreasonable Reasonable algorithms May be usable depending upon the input size Unreasonable algorithms Are impractical and useful to theorists Demonstrate need for approximate solutions Remember were dealing with large N (input size)

Slide 46

Two Categories of Algorithms 2 4 8 16 32 64 128 256 512 1024 Size of Input (N) 10 35 10 30 10 25 10 20 10 15 trillion billion million 1000 100 10 N N5N5 2N2N N Unreasonable Dont Care! Reasonable Runtime

Slide 47

Summary Reasonable algorithms feature polynomial factors in their O() and may be usable depending upon input size. Unreasonable algorithms feature exponential factors in their O() and have no practical utility.

Slide 48

Questions?

Slide 49

Using O() Analysis in Design

Slide 50

Air Traffic Control Coast, add, delete Conflict Alert

Slide 51

Problem Statement What data structure should be used to store the aircraft records for this system? Normal operations conducted are: Data Entry: adding new aircraft entering the area Radar Update: input from the antenna Coast: global traversal to verify that all aircraft have been updated [coast for 5 cycles, then drop] Query: controller requesting data about a specific aircraft by location Conflict Analysis: make sure no two aircraft are too close together

Slide 52

Air Traffic Control System ProgramAlgorithmFreq 1. Data Entry / ExitInsert15 2. Radar Data UpdateN*Search12 3. Coast / DropTraverse60 4. QuerySearch 1 5. Conflict AnalysisTraverse*Search12

Slide 53

Questions?

Slide 54

Concurrent Systems

Slide 55

Sequential Processing All of the algorithms weve seen so far are sequential: They have one thread of execution One step follows another in sequence One processor is all that is needed to run the algorithm

Slide 56

A Non-sequential Example Consider a house with a burglar alarm system. The system continually monitors: The front door The back door The sliding glass door The door to the deck The kitchen windows The living room windows The bedroom windows The burglar alarm is watching all of these at once (at the same time).

Slide 57

Another Non-sequential Example Your car has an onboard digital dashboard that simultaneously: Calculates how fast youre going and displays it on the speedometer Checks your oil level Checks your fuel level and calculates consumption Monitors the heat of the engine and turns on a light if it is too hot Monitors your alternator to make sure it is charging your battery

Slide 58

Concurrent Systems A system in which: Multiple tasks can be executed at the same time The tasks may be duplicates of each other, or distinct tasks The overall time to perform the series of tasks is reduced

Slide 59

Advantages of Concurrency Concurrent processes can reduce duplication in code. The overall runtime of the algorithm can be significantly reduced. More real-world problems can be solved than with sequential algorithms alone. Redundancy can make systems more reliable.

Slide 60

Disadvantages of Concurrency Runtime is not always reduced, so careful planning is required Concurrent algorithms can be more complex than sequential algorithms Shared data can be corrupted Communications between tasks is needed

Slide 61

Achieving Concurrency CPU 1 CPU 2 Memory bus Many computers today have more than one processor (multiprocessor machines)

Slide 62

Achieving Concurrency CPU task 1 task 2 task 3 ZZZZ Concurrency can also be achieved on a computer with only one processor: The computer juggles jobs, swapping its attention to each in turn Time slicing allows many users to get CPU resources Tasks may be suspended while they wait for something, such as device I/O

Slide 63

Concurrency vs. Parallelism Concurrency is the execution of multiple tasks at the same time, regardless of the number of processors. Parallelism is the execution of multiple processors on the same task.

Slide 64

Types of Concurrent Systems Multiprogramming Multiprocessing Multitasking Distributed Systems

Slide 65

Multiprogramming Share a single CPU among many users or tasks. May have a time-shared algorithm or a priority algorithm for determining which task to run next Give the illusion of simultaneous processing through rapid swapping of tasks (interleaving).

Slide 66

Multiprogramming Memory User 1 User 2 CPU User1 User2

Slide 67

Multiprogramming 1 2 3 4 1 234 CPUs Tasks/Users

Slide 68

Multiprocessing Executes multiple tasks at the same time Uses multiple processors to accomplish the tasks Each processor may also timeshare among several tasks Has a shared memory that is used by all the tasks

Slide 69

Multiprocessing Memory User 1: Task1 User 1: Task2 User 2: Task1 CPU User1 User2 CPU

Slide 70

Multiprocessing 1 2 3 4 1 234 CPUs Tasks/Users Shared Memory

Slide 71

Multitasking A single user can have multiple tasks running at the same time. Can be done with one or more processors. Used to be rare and for only expensive multiprocessing systems, but now most modern operating systems can do it.

Slide 72

Multitasking Memory User 1: Task1 User 1: Task2 User 1: Task3 CPU User1

Slide 73

Multitasking 1 2 3 4 1 234 CPUs Tasks Single User

Slide 74

Distributed Systems Central Bank ATM Buford ATM Perimeter ATM Student Ctr ATM North Ave Multiple computers working together with no central program in charge.

Slide 75

Distributed Systems Advantages: No bottlenecks from sharing processors No central point of failure Processing can be localized for efficiency Disadvantages: Complexity Communication overhead Distributed control

Slide 76

Questions?

Slide 77

Parallelism

Slide 78

Using multiple processors to solve a single task. Involves: Breaking the task into meaningful pieces Doing the work on many processors Coordinating and putting the pieces back together.

Slide 79

Parallelism CPU Memory Network Interface

Slide 80

Parallelism 1 2 3 4 1 234 CPUs Tasks

Slide 81

Pipeline Processing Repeating a sequence of operations or pieces of a task. Allocating each piece to a separate processor and chaining them together produces a pipeline, completing tasks faster. ABCD input output

Slide 82

Example Suppose you have a choice between a washer and a dryer each having a 30 minutes cycle or A washer/dryer with a one hour cycle The correct answer depends on how much work you have to do.

Slide 83

One Load washdry combo Transfer Overhead

Slide 84

Three Loads washdry combo wash dry combo

Slide 85

Examples of Pipelined Tasks Automobile manufacturing Instruction processing within a computer 154 32 1 23 4 5 1 23 4 5 1 2 3 4 5 A B C D 12345670 time

Slide 86

Task Queues P1P2P3Pn Super Task Queue A supervisor processor maintains a queue of tasks to be performed in shared memory. Each processor queries the queue, dequeues the next task and performs it. Task execution may involve adding more tasks to the task queue.

Slide 87

Parallelizing Algorithms How much gain can we get from parallelizing an algorithm?

Slide 88

Parallel Bubblesort 9387746557453327 9387746557453327 9387746557453327 We can use N/2 processors to do all the comparisons at once, flopping the pair-wise comparisons.

Slide 89

Runtime of Parallel Bubblesort 93877465574533273 93877465574533274 93877465574533275 93877465574533276 93877465574533277 93877465574533278

Slide 90

Completion Time of Bubblesort Sequential bubblesort finishes in N 2 time. Parallel bubblesort finishes in N time. Bubble Sort parallel O(N 2 ) O(N)

Slide 91

Product Complexity Got done in O(N) time, better than O(N 2 ) Each time chunk does O(N) work There are N time chunks. Thus, the amount of work is still O(N 2 ) Product complexity is the amount of work per time chunk multiplied by the number of time chunks the total work done.

Slide 92

Ceiling of Improvement Parallelization can reduce time, but it cannot reduce work. The product complexity cannot change or improve. How much improvement can parallelization provide? Given an O(NLogN) algorithm and Log N processors, the algorithm will take at least O(?) time. Given an O(N 3 ) algorithm and N processors, the algorithm will take at least O(?) time. O(N) time. O(N 2 ) time.

Slide 93

Number of Processors Processors are limited by hardware. Typically, the number of processors is a power of 2 Usually: The number of processors is a constant factor, 2 K Conceivably: Networked computers joined as needed (ala Borg?).

Slide 94

Adding Processors A program on one processor Runs in X time Adding another processor Runs in no more than X/2 time Realistically, it will run in X/2 + time because of overhead At some point, adding processors will not help and could degrade performance.

Slide 95

Overhead of Parallelization Parallelization is not free. Processors must be controlled and coordinated. We need a way to govern which processor does what work; this involves extra work. Often the program must be written in a special programming language for parallel systems. Often, a parallelized program for one machine (with, say, 2 K processors) doesnt work on other machines (with, say, 2 L processors).

Slide 96

What We Know about Tasks Relatively isolated units of computation Should be roughly equal in duration Duration of the unit of work must be much greater than overhead time Policy decisions and coordination required for shared data Simpler algorithm are the easiest to parallelize

Slide 97

Questions?

Slide 98

More?

Slide 99

Matrix Multiplication

Slide 100

Inner Product Procedure Procedure inner_prod(a, b, c isoftype in/out Matrix, i, j isoftype in Num) // Compute inner product of a[i][*] and b[*][j] Sum isoftype Num k isoftype Num Sum

i N) server