A5: Search Accelerated

Search Accelerated

In this assignment you will finish building the search engine that you began in A4.

What you’ll do: Implement red-black trees, and test the performance of your engine.

Table of contents:

Step 1: Team Maintenance

Your A5 team will be the same as your A4 team. At your first team meeting about A5, please assign roles for this assignment.

After this assignment is due, you will submit a performance evaluation of your teammates (covering A2 through A5) to the professor, and those evaluations will become part of team members’ grades.

Step 2: Explore the Starter Code

There is a makefile provided with the usual targets: build, test, check, finalcheck, zip, docs, and clean. As usual, you may not change the names and types in the provided interfaces, and your submission must pass make check. If you’re ever in doubt about whether a change is permitted or not, just run make check: it will tell you whether you have changed the interface in a prohibited way. There is a new target, make perf, which is related to performance testing and is discussed in that step, below.

You will want to replace the blank implementations in the A5 starter code with your implementations from A4. Here’s how to do that:

FYI: we also provide a Bisect target, as in A4, but the make test and make bisect targets are factored differently in this assignment than in A4. In A4, make test linked in the Bisect tool, which meant all your tests were being instrumented whether you cared about code coverage or not. In A5, make test does not link Bisect, but make bisect does. The reason for this change is the performance testing that A5 asks you to do: Bisect greatly slows down your code, because of the instrumentation it adds, and we don’t want you to have artificially slow code during performance testing.

Step 3: Finish A4

If there’s anything your team didn’t finish in A4, begin A5 by finishing it now. That includes fixing any faults that might have been discovered by the A4 autograder. Make sure to add unit tests for any such faults. Also make sure that you have built out your test suite fairly well, because you’re going to need good tests when it comes to red-black trees.

To be clear, your association list dictionary and list-based search engine are a required component of A5. Keep those working as you implement everything else.

But there is one thing you may omit from A4: Bisect. On A5 you are not required to achieve any particular code coverage, or to submit a Bisect report. On the other hand, you are strongly encouraged to make use of Bisect as you see fit to help you improve your A5 test suite.

Step 4: Red-Black Trees

The primary performance bottleneck in your search engine so far is the dictionary data structure. In particular,

Let’s improve on that by implementing dictionaries with red-black trees. They offer logarithmic time find, insert, and remove operations. It’s the faster find operation that will dramatically improve your search engine. Each node in your red-black tree will store a single binding from a key to a value, and the tree will be ordered according to the keys. That is, the binary search tree (BST) invariant, which red-black trees must satisfy, will apply to the keys: all nodes in a left subtree must have a key less than the key at the parent, and all nodes in a right subtree must have a key greater than the key at the parent.

Implement TreeDictionary by finishing the code in treeDictionary.ml. But, you should omit remove for now. It is considerably more challenging, hence will be part of the Excellent scope. If your engine.ml happened to use remove, you should replace that code now with another algorithm that doesn’t require remove. Otherwise you could end up not having a working tree-based search engine, thus not even getting the points for the Satisfactory scope.

The example in exampleDictionary.ml has been updated to use TreeDictionary, in case you would again like guidance on how to create a dictionary in utop.

When you’ve finished implementing the tree dictionary, the TreeEngine provided in the starter code will make use of the Engine and DictionarySet modules you already completed in A4 to produce a fast search engine. So, the only new piece of implementation for this step is treeDictionary.ml, and the unit tests that you implement for it in test.ml.

Tip 1. Consider building out your red-black tree in two phases:

Tip 2. As you test your tree dictionary and engine, you should be able to re-use your list-based tests. Consider not copy-pasting the tests, but instead using a functor to create a test suite. The input to that functor would be the structure that you want to test, and the output would be a structure that contains a list of OUnit tests. You can then pass that list to the OUnit test runner.

Tip 3. To use TDD, begin by commenting out all your tests you already created for lists. Then uncomment the tests for whatever operation you want to implement next. Implement it for trees. Make sure your trees and lists pass those tests. Repeat until all tests are uncommented.

Tip 4. Use rep_ok to help you. Implement it first. Make sure all your tree operations check that rep_ok holds of inputs and outputs. This strategy will make you a more productive programmer: you will quickly discover your faults and fix them, rather than having to live with mysterious or unknown faults. After all your tests are passing, feel free to comment out your (probably expensive) implementation of rep_ok and replace it with something like let rep_ok x = x, or even

let debug = false
let rep_ok x = 
  if debug then (* your actual implementation *) 
  else x

Then later you could quickly re-enable rep_ok by changing debug to true, if you suspected there was new fault you needed to diagnose.

Note: Unlike tail recursion, which requires only constant stack space, the red-black tree algorithms require logarithmic stack space. That’s okay. You shouldn’t attempt to rewrite the tree algorithms to use constant stack space.

This is the stopping point for a satisfactory solution.

Step 5: Performance Testing

The reason you just implemented red-black trees was to get better performance. Probably you’ve already noticed that your tree-based tests run faster, especially on large directories. But how do you know that you actually achieved a logarithmic time implementation? Looking at individual running times doesn’t answer that question. Instead, you need to look at how the data structure’s operations scale as the number of items it contains grows larger.

Here are two graphs that show the scaling of list sets vs. tree sets (i.e., a DictionarySet made from ListDictionary vs. a DictionarySet made from TreeDictionary):

List Performance Tree Performance

The first graph shows the performance of list sets; the second, tree sets. Note the axes of both graphs carefully: the y-axis is execution time in seconds, and is around 0-7 seconds for both graphs; whereas the x-axis is the number of elements in the set, and is quite different between the two graphs. For lists, the number of elements reaches only 10,000. For trees, it reaches 1,000,000.

The execution time being measured is the time it takes to do the following operations, where \(n\) is the number of elements:

(Note: that workload will change a little for the Excellent scope.)

Note also that there are two lines shown in each graph. These represent two different workloads: ascending vs. random. In the ascending workload, the elements are first sorted before being inserted. In the random workload, they remain in a random order. As we saw in lecture, the difference between a random workload and an ascending workload matters, because an unbalanced tree data structure would do poorly on the latter.

And every data point reported in the graph is actually the median of three repetitions of the experiment, in an effort to reduce noise.

The list graph clearly shows that the implementation is scaling quadratically: there are \(O(n)\) operations being performed, each of which is \(O(n)\), for a total of \(O(n^2)\) work. But the tree graph is much better: each operation is only \(O(\log n)\), so the total work is \(O(n \log n)\). That’s why the graph looks nearly linear, but has a slight upward curve. The list implementation does slightly worse for the ascending workload; that’s probably because the particular list implementation used in producing that graph keeps the list in sorted order, which causes every insert to take a full \(n\) operations to walk down the whole list. The tree implementation does worse for the ascending random workload; that’s probably because it triggers more rebalancing operations. But in both cases, the two workloads are fairly similar in execution time.

We produced those graphs by writing OCaml code that performs the workloads and does the timing. That OCaml code wrote out CSV files with the results. We then used a graphing utility, gnuplot, to create graphs from the CSVs.

Your task: Do that same analysis for your own implementation. That is, produce two graphs, each of which shows those two workloads, for the same maximum values on the x axis. Then explain the results that you see. Submit as part of your zipfile a file named analysis.pdf that contains your graphs and your explanation.

What doesn’t matter: The absolute numbers on your y-axis don’t matter. If your implementation happens to take more like 5 seconds or 10 seconds or 30 seconds, that’s fine. What we’re interested in here is the asymptotic behavior, not the constants. After all your machine might just be faster or slower than ours. Also, you don’t have to completely automate your experiments like we did, or use CSVs, or use gnuplot. But you would be wise to automate as much as you can, because undoubtedly you will want to run the experiments more than once.

What does matter: We want to see quadratic performance on lists and linearithmic—i.e., \(O(n \log n)\)— on trees, for both workloads. If you aren’t getting those running times, then you have a performance bug, and you need to fix it.

What matters even more: Fabrication and falsification of data is unethical. Science rests upon the honesty of experimenters. If you were to make up or modify the data that you present in your graphs, then you would be in violation of the Code of Academic Integrity. (And if you were a professional researcher, you could get fired.) So just don’t do it. Even if you can’t resolve your performance bugs, it’s better to acknowledge that in your analysis than to lie about your experiments.

Starter code: In performanceTest.ml we provide some starter code to help you, including generation of random and sorted lists, as well as a function to time the execution time of other functions. We also provide the two CSV files for our graphs above and the gnuplot script that produces the graphs from them, in case you want to attempt the same automation. (If you want to use gnuplot you will need to install it through your Linux package manager, e.g., sudo apt-get install gnuplot or sudo port install gnuplot.) The make perf target runs those. It also runs the OCaml native compiler instead of the bytecode compiler. Native mode produces faster code, which means your experiments will run more quickly. Using native mode is not required but is likely to be helpful.

This is the stopping point for a good solution.

Step 6: Remove

All that remains is to implement the remove operation for red-black trees. This will be perhaps the most challenging Excellent scope yet!

The algorithm you should use is that defined in Section 2 of this short paper by Germane and Might. Although the code in that paper is written in Racket and Haskell, two other functional languages, the algorithm is fully described in English and pictures. So the challenge will be understanding and implementing the algorithm: a challenge that you will likely face in the future in other programming tasks! So, this should be good practice.

There are just two places we found any ambiguity in their description of the algorithm:

  1. For their Figures 6, 7, and 9: the double black nodes shown in those could in fact be double black leaves.

  2. When they say the redden operation changes the root node to be red “if possible,” they mean: if the root is black with two black children, which themselves are black nodes or (black) leaves. (Update 10/15/18: this text was incorrect: or if the root is a (black) leaf.) And when they say the blacken operation changes the root node to be black “if necessary,” they mean: if the root is red with at least one red child.

Also, note carefully when the paper mentions the “reflections” of trees at the bottom of p. 426: your code will need to account for those reflections.

After you have implemented remove (which took us around 100 lines of code, including modifying the representation type to account for double black nodes and leaves, and therefore modifying some of pattern matches in the other operations), make sure that you have tested it. Especially, make sure that you check for the representation invariant always holding.

Finally, update your performance analysis to include remove in the workload:

You should still see linearithmic performance.

Scope

Here’s what we consider a satisfactory, good, and excellent solution:

Documentation and Testing

As in A2 and A3, the graders will look at the documentation produced by make docs. As in A4, the graders will not run make test on your submission, nor do we expect you to submit test directories; instead, the graders will read your test.ml.

Submission

Make sure your team’s NetIDs are in authors.mli, and set the hours_worked variable at the end of authors.ml.

Run make zip to construct the ZIP file you need to submit on CMS. Any mal-constructed ZIP files will receive a penalty of 20 points. The size of the the files is limited in CMS to 1 MB. Please stay within the size allotted, and do not submit large (or even any) test directories. Make sure your analysis.pdf file, including your graphs and explanation, are in your zip.

Submit your zipfile on CMS. Double-check before the deadline that you have submitted the intended version of your file.

Congratulations! Your search has come to an end.


Acknowledgement: Adapted from Prof. Greg Morrisett, Dean of Cornell CIS.