EECS 380 Homework 3

Winter Semester '01

Assigned Feb 15th 2001.
Due Monday, March 12th, 6pm, EECS 380 box in room 2420 EECS

You are to turn this in with all parts stapled together. Please use two-sided printing to reduce the amount of paper in homeworks, if you can (graders are having trouble carrying all homeworks!). Make sure to mention the following on the first page: your name, uniquename and the number/time of your discussion section.


  1. Racing STL's sort() versus the qsort() function from stdlibc
  2. Improving the worst-case time of the quicksort algorithm
  3. Incremental sorting
  4. Visualization of bubble sort and shaker sort

1. Racing STL's sort() versus qsort() from stdlibc

You need to compare the runtime of sort() from STL to qsort() (provided in the C standard library) on large arrays of large integers. You need to conclude which one is faster on average and by how much (a multiplicative factor).

A sample call to STL's sort is given in the STL's documentation. A sample call to qsort is given in Sedgewick on p. 303. Additionally, all functions from the C standard library have man pages (try man qsort, a caveat in this man page is discussed in the newsgourp).

You need to use the following comparison methodology:

  1. For a given N, populate an array with N randomly generated large integers. For this use the srand() and rand() calls (or srandom() and random()). Type man rand and man random for details on those two (here's an optional journal article on rand() and random()). It is recommended that you print your random numbers when testing the code to make sure they "look random".

  2. You will need two copies of the array --- one for sort and the other for qsort. Run sort and qsort once, and measure the cpu time for each. Make sure that N is large enough to give reliable readings without repeated runs. You must compile your code with the compiler flag -O3

  3. Change the random seed (otherwise, you will always generate the same "random" numbers), repeat the experiment for the same N, compute the average runtime over all experiments done so far for the same N. You can stop when the average stabilizes to within 10%.

  4. For a given N, you need to figure out which sorting call is faster and by how much on average (by which factor). Use three values N>10000, each at least 2x apart from each other.
You do not need to plot anything in this question. You need to produce 9 numbers: for each of the three N values you select, you will have two averaged runtimes and a comparison factor (the ratio between the two runtimes). The final answer in English should say whether qsort() or sort() is recommended when speed is all that matters. You should compile with -O3, and are very likely to get a wrong answer otherwise.

Feel free to explain your results, but explanations will not be graded.

2. Improving the worst-case time of qsort

Using random inputs and the averaging methodology from part1, empirically show that the STL's nth_element function runs in linear time on average when it is used to find the median (i.e., n is half the size of the container). For this, build a plot of its average run time as a function of the input size for at least 6 (significantly different) values of N. Every datapoint needs to be averaged as explained in part1, but you don't need to compare it to anything.

Assume that you are given a function median() (with whatever arguments you wish) that computes the median in worst-case linear time (STL's implementation apparently does not). A very non-trivial algorithm for this is explained in detail in the CLR book (pp.189-196), but you are not required to read that --- just assume that you can use its implementation.

  1. Given the example of worst-case input for the original quick sort algorithm described in the Sedgwick book (page 321); explain why the median-of-three pivoting and choosing a pivot at random, each, do not improve the worst-case big-theta complexity of qsort.

  2. Consider modifying the original qsort algorithm using linear-time median selection so that the modified algorithm will have better worst-case complexity. Consider two cases: (i) each element in the array can only repeat O(1) times, (ii) the general case. You need to provide a proof of its worst-case complexity (it can refer to the proof of worst case complexity of qsort and simply analyze modifications).
In practice, linear-time median does a lot of work, and "improved" qsort() may not be competitive with original qsort on average, despite better asymptotic worst-case performance and asymptotically equivalent average-case performance.

3. Incremental sorting

Generate large random almost-sorted ranges. To do this you will start with a randomly generated array of integers from part 1 and sort such an array; this will give you a large random sorted range. Now let's spoil it. Pick two indices from 0 to N-1 at random ( idx=rand() % N; ) and swap the values. In some application this can correspond to a perturbation of an already-sorted range, so we need to bring it back to sorted order.

4. Visualization of bubble sort and shaker sort

  1. You need to have a working implementation of bubble sort for arrays (can paste from Sedgewick). Shaker sort is a variant of bubble sort in which passes are performed left-to-right and right-to-left in turns, rather than in one direction as in bubble sort (why would this be useful?). You need a shaker sort implementation as well. Both implementations should terminate once the range is sorted (rather than perform "dummy" passes).

  2. Using the index-sorting implementation in the handout, write a function plot() that takes a range and saves its sorted index so that the index can be plotted by gnuplot (i.e., two unsigned's per line).

  3. Use N=200 for your experiments, use your student ID as random seed. Generate an array of N random numbers as in part 1. Find out how many passes bubble sort and shaker sort perform on this array. Modify the codes of bubble sort and shaker sort so that plot() is called half way through (e.g., if there are 90 passes total, plot() should be called at pass 45). You need to produce two pictures. They will probably look similar to Fig 6.7 in Sedgewick.

Turn-in check-list

   Part 1: * Source code (one file). 
           * A table with 9 numbers.
	   * A recommendation in English as to whether
	     sort() or qsort() should be used when speed matters.

   Part 2: * A plot showing that STL's median selection runs in linear time 
             on average. 
	   * A description of worst-case inputs for qsort. 
	   * An explanation of why both median-of-three and random pivoting
	     do not change worst-case big-theta.  
	   * A [theoretical] proof of the worst-case run time when using 
             the median as the pivot.

               [You do not need to implement this modified quicksort]

   Part 3: * Source code (one file). 
           * One plot showing that your algorithm has linear average-case 
	     complexity on the almost-sorted ranges considered (plot as in
	     part 2)
	   * A comparison between your algorithm and the faster of
	     sort()/sqort() --- table as in part 1.

   Part 4: * Source code for bubble and shaker sort.
	   * An explanation why it makes sense to perform bubble-passes in
	     both directions. 
	   * Two plots.  (Plots produced by different students
	     must be different.  If plots submitted by different students are
	     the same, both homeworks may be disqualified (0 score for the
	     whole homework)).