Week 12.pptx

CS 177 Week 12: Running Time and Performance How long does a program take to run? Most of the problems you have written in this class run in a few seconds or less Some kinds of programs can take much longer: Chess algorithms (Deep Blue) Routing and scheduling algorithms (Traveling Sales Person, Purdue classroom scheduling) Medical imaging (massive amounts of data) Graphics processing (Pixar: Toy Story, Incredibles) How can you measure the amount of time a program takes? You can time the program What if it takes 1000 years to run? You can look at the program What are you looking for? Does the input matter? Sorting 100,000 numbers must take longer than sorting 10 numbers, right? Algorithms A sequence of steps used to solve a problem In CS, we take an algorithm and convert it into code -- Java in this class The study of algorithms is intended to: Make sure that the code you write gets the right answer Make sure that the code runs as quickly as possible Make sure that the code uses a reasonable amount of memory Algorithmic example: search You want to find the definition of the word ?rheostat? in the dictionary The Oxford English Dictionary has 301,100 main entries What if you search for the word ?rheostat? by starting at the beginning and checking each word to see if it is the one we are interested in Given a random word, you?ll have to check about 150,550 entries before you find it Alternatively, you could jump to the correct letter of the alphabet, saving tons of time Searching for ?zither? takes forever otherwise Algorithmic example: Summation Write a program to sum up the numbers between 1 and n Easy, right? Let us recall that a faster way might be to use the summation equation: Memory usage Memory usage is a problem too If you run out of memory, your program can crash Memory usage can have serious performance consequences too Memory Remember, there are multiple levels of memory on a computer Each next level is on the order of 100 times larger and 100 times slower 100X 100X 100X 100X Size Speed Memory speed If you can do a lot of number crunching without leaving cache, that will be very fast If you have to fetch data from RAM, that will slow things down If you have to read and write data to the hard drive (unavoidable with large pieces of data like digital movies), you will slow things down a lot Measuring memory Memory can be easier to estimate than running time Depending on your input, you will allocate a certain number of objects, arrays, and primitive data types It is possible to count the storage for each item allocated A reference to an object or an array costs an additional 4 bytes on top of the size of the object What?s a graph? A graph is a set of nodes and edges It?s a really simple way to represent almost any kind of relationship The numbers on the edges could be distances, costs, time, whatever 5 3 4 16 A B D C E 2 8 6 Traveling Salesperson Problem There?s a classic CS problem called the Traveling Salesperson Problem (TSP) Given a graph, find the shortest path that visits every node and comes back to the starting point Like a lazy traveling salesperson who wants to visit all possible clients and then return home as soon as possible How would you solve the TSP? Strategies: Always visit the nearest neighbor next Randomized approaches Try every possible combination How can you tell what?s the best solution? Closest doesn?t always work We are tempted to always take the closest neighbor, but there are pitfalls Trying every combination is brutal In a completely connected graph, we can try any sequence of nodes If there are n nodes, there are (n ? 1)! tours For 30 cities, 29! = 8841761993739701954543616000000 If we can check 1,000,000,000 tours in one second, it will only take about 20,000 times the age of the universe to check them all We will (eventually) get the best answer! No one knows the best solution This and similar problems are useful problems to solve for people who need to plan paths UPS Military commanders Pizza delivery people Internet routing algorithms No one knows a solution that always works in a reasonable amount of time We have been working on it for decades Fabulous cash prizes The problem is so hard that you can win money if you solve it Clay Mathematics Institute has offered a $1,000,000 prize You can do one of two things to collect it: Find an efficient solution to the general problem Prove that it is impossible to find an efficient solution in some cases What does efficient mean? How do we measure algorithm running time and memory usage in general? We want to take into account the size of the problem (TSP with 4 cities is certainly easier than TSP with 4,000 cities) Some algorithms might work well in certain cases and badly in others How do we come up with a measuring scheme? Classification Ideally, we want to classify both problems and algorithms We want to figure out which problems take a really long time to solve We want to figure out how we can compare one algorithm to another We want to prove that some algorithm is the best possible algorithm to solve a particular problem Big-Oh Notation Computer scientists use Big-Oh notation to give an indication of how much time (or memory) a program will take Big-Oh notation is: Asymptotic Focuses on the behavior as the size of input gets large Worst-Case Focuses on the absolute longest the program can take Running time example Add up the operations done by the following code: Initialization: 1 operation Loop: 1 initialization + n checks + n increments + n additions to sum = 3n + 1 Output: 1 operation Total: 3n + 3 int sum = 0; for( int i = 0; i < n; i++ ) sum += i; System.out.println(?Sum: ? + sum); Running time function We could express the time taken by the code on the previous slide as a function of n: f(n) = 3n + 3 This approach has a number of problems: We assumed that each line takes 1 time unit to accomplish, but the output line takes much longer than an integer operation This program is 4 lines long, a longer program is going to have a very messy running time function We can get nit picky about details: Does sum += i; take one operation or two if we count the addition and the store separately? Inadequacy of running time function In short, this way of getting a running time function is almost useless because: It cannot be used to give us an idea of how long the program really runs in seconds It is complex and unwieldy The most important thing about the analysis of the code that we did is learning that the growth of the function should be linear (3n+3) A general description of how the running time grows as the input grows would be useful Big Oh notation Enter Big Oh notation Big Oh simplifies a complicated running time function into a simple statement about its worst case growth rate All constant coefficients are ignored All low order terms are ignored 3n + 3 is O(n) Big Oh is a statement that a particular running time is no worse than some function, with appropriate constants Big Oh simplification examples 147n3 + 2n2 + 5n + 12083 is O(n3) 15n2 + 6n + 7log n + 145 is O(n2) log n + 87829 is O(log n) Note: In CS, we use log2 unless stated otherwise Mulitiplication by hand How long does it take to do multiplication by hand? 123 x 456 738 615 492__ 56088 Let?s assume that the length of the numbers is n digits (n multiplications + n carries) x n digits + (n + 1 digits) x n additions = 3n2 + n Running time: O(n2) Finding the largest element in an array How do we find the largest element in an array? Running time: O(n) if n is the length of the array What if the array is sorted in ascending order? Running time: O(1) int largest = array[0]; for( int i = 1; i < array.length; i++ ) if( array[i] > largest ) largest = array[i]; System.out.println(?Largest: ? + largest); System.out.println(?Largest: ? + array[array.length-1]); Multiplying matrices Given two n x n matrices A and B, the code to multiply them is: Running time: O(n3) Is there a faster way to multiply matrices? Yes, but it?s complicated and has other problems double[][] c = new double[n][n]; for( int i = 0; i < n; i++ ) for( int j = 0; j < n; j++ ) for( int k = 0; k < n; k++ ) c[i][j] += a[i][k]*b[k][j]; Bubble sort Here is some code that sorts an array in ascending order What is its running time? Running time: O(n2) for( int i = 0; i < array.length; i++ ) for( int j = 0; j < array.length - 1; j++ ) if( array[j] > array[j + 1] ) { int temp = array[j]; array[j] = array[j + 1]; array[j + 1] = temp; } Rules of thumb There is nothing better than constant time (O(1)) Logarithmic time (O(log n)means that the problem can become much larger and only take a little longer Linear time (O(n))means that time grows with the problem Quadratic time (O(n2)) means that expanding the problem size significantly could make it impractical Cubic time (O(n3)) is about the reasonable maximum if we expect the problem to grow Exponential and factorial time mean that we cannot solve anything but the most trivial problem instances Empirical techniques Looking at the code and determining the order of complexity is one technique Sometimes this only captures worst-case behavior The code can be complicated and difficult to understand Another possibility is running experiments to see how the running time changes as the problem size increases Doubling hypothesis The doubling hypothesis states that it is often possible to determine the order of the running time of a program by progressively doubling the input size and measuring the changes in time Finding the largest element in an array Let?s use the doubling hypothesis to test the running time of the program that finds the largest element in an unsorted array Looks like a decent indication of O(n) Size Time Ratio 1024 0.035829 ms - 2048 0.071727 ms 2.00 4096 0.188013 ms 2.62 8192 0.278877 ms 1.48 Multiplying matrices Let?s check matrix multiplication A factor of 8 is expected for O(n3) Not too bad Size Time Ratio 128 39.786341 ms - 256 229.902823 ms 5.78 512 2945.396691 ms 12.81 1024 25957.367417 ms 8.81 Bubble sort Let?s try bubble sort Success! O(n2) is supposed to give a ratio of 4 We should have done more trials to do this scientifically Size Time Ratio 1024 8.648306 ms - 2048 28.331604 ms 3.28 4096 108.931582 ms 3.84 8192 404.451683 ms 3.71