HWcpp

See the lecture

Overview

We will be solving a problem where we are given a list of documents. Each document consists of multiple words separated by space. We will assign each document an ID 0,1,... in the order they appear on input. We want to implement an indexing data structure that allows the following operations:

• Function add_document adds a document to the index and returns its ID (the next available ID). A document is a list of words (strings).
• Function search finds all documents that contain a given word, and returns the list of their IDs in increasing order.
• Function multi_search gets a list of words (W1, W2, ...) and finds all documents that contain all of the words (so each document has to contain all words from the query). It returns the list of document IDs in increasing order.

Note that this functionality could be a base for a very simple text search engine. We could extend it with the option for storing and printing whole documents containing given keywords of interest, but we will not do this today.

In this homework, you are given a Python program that implements these functions using a structure called inverted index.

• The structure consists of a dictionary in which keys are the words. For each word we keep a list of document IDs in the increasing order.
• Function search simply returns the list from the dictionary.
• Function multi_search uses search to return a list of IDs for each input word. It maintains an intersection of lists seen so far and computes pairwise intersection of the new list and previous intersection.
• Intersection of two sorted lists of IDs is done by a procedure similar to merging two sorted lists in Merge Sort. One integer index is moved along each list so that we always move forward the one that points to a smaller value. If both indices point to the same value, we found an element from the intersection which is placed into the result. As soon as one index reaches the end of the list, we are finished.

The overall program doc_finder.py requires two command-line arguments. The first is a file with documents, one document per line. Words in the document are separated by spaces. The second argument is a file with queries, one query per line. If the line contains a single word, it is passed to search, otherwise to multi_search. For each query, the program prints the IDs of the found documents separated by spaces. If no documents were found, it prints a dash.

Here is an example of tiny debugging inputs and outputs:

==> tiny_documents.txt <==
this is document zero about frog
this is document one about dog and it is longer

==> tiny_queries.txt <==
dog
cat
is
is frog
dog frog
cat frog
is this

==> tiny_answers.txt <==
1
-
0 1
0
-
-
0 1

Files

Tiny debugging input and correct output:

• /tasks/cpp/tiny_documents.txt a file with documents
• /tasks/cpp/tiny_queries.txt a file with queries (both single and multi-word)
• /tasks/cpp/tiny_answers.txt correct answers for the queries

Larger inputs:

• /tasks/cpp/documents.txt a file with documents
• /tasks/cpp/queries.txt a file with single-word queries
• /tasks/cpp/multiqueries.txt a file with multi-word queries

Program in Python:

• /tasks/cpp/doc_finder.py

You can run the program as follows:

# run the program on the tiny inputs
python3 /tasks/cpp/doc_finder.py /tasks/cpp/tiny_documents.txt /tasks/cpp/tiny_queries.txt > test_answers.txt
# check whether the answer is correct
diff /tasks/cpp/tiny_answers.txt test_answers.txt

Task A

The functions listed above are implemented in class DocumentIndex in file doc_finder.py. Reimplement this class in C++ and use it in Python.

• Your Python script should be named doc_finder_cpp.py. Function main from doc_finder.py can remain almost the same as in doc_finder.py except for a slightly different initialization of index variable using the C++ class. Omit the Python implementation of the class for the file.
• The C++ class should be placed in file DocumentIndex.cc
• Use the same algorithm in C++ as in Python. We recommend using STL structures unorderd_map and vector.
• Submit files doc_finder_cpp.py and DocumentIndex.cc.
• Your implementation should successfully compile and run on the tiny inputs and produce correct output (empty diff) using the commands below:

# compile C++
c++ -O3 -Wall -shared -std=c++11 -fPIC $(python3 -m pybind11 --includes) DocumentIndex.cc -o DocumentIndex$(python3-config --extension-suffix)
# run the program on the tiny inputs
python3 doc_finder_cpp.py /tasks/cpp/tiny_documents.txt /tasks/cpp/tiny_queries.txt > test_answer.txt
# check whether the answer is correct
diff /tasks/cpp/tiny_answers.txt test_answer.txt

Task B

• Compare the running time and memory required by doc_finder.py and doc_finder_cpp.py.
• Use the provided list of documents /tasks/cpp/documents.txt and consider three files with queries:
  • Empty file with queries to measure how long it takes to build the index
  • /tasks/cpp/queries.txt to measure speed of search
  • /tasks/cpp/multiqueries.txt to measure the speed of multi-search
• You can measure time and memory using /usr/bin/time -v, such as:

/usr/bin/time -v python3 doc_finder.py /tasks/cpp/documents.txt /tasks/cpp/queries.txt > tmp-output.txt > time-measurement.txt

• Interesting rows from time measurements are Elapsed (wall clock) time, Maximum resident set size; also check Percent of CPU this job got
• Run each program several times (at least 3x) to see how the time and memory fluctuate.
• Report the results in the protocol (in some easy to read form). Consider using summary statics such as mean and standard deviation. Optionally you can show the results in a plot or a table placed in a png or pdf file. Do not forget to discussion of the results.
• Also discuss your conclusions from the measured values in the protocol.
• Submit the protocol.

Notes:

• It might be a good idea to write some script / makefile for running programs and parsing the time and memory values. If you do so, submit your script as well.
• You can run the experiments on your computer or on vyuka. If you run on vyuka, make sure the server is not too busy (you should get a high % of CPU, close to 100%).

Task C

In this task you choose one the following two options for extending your work from tasks A and B.

Option 1

Try a different implementation of the data structure in both Python and C++. One example is using sets instead of lists of document IDs for each word and then using set intersection provided by language libraries instead of manual merge algorithm. You can also come up with a different approach, but try to keep it sensible. Compare the time and memory of your new implementations to the results from Task B.

• Submit your implementations.
• In the protocol describe the approach you chose in your implementation.
• In the protocol, also give the results of the time and memory measurements and discuss your results (compare the two versions but also compare them with the previous two).

Option 2

Do a more extensive profiling of the two implementations you already have. Here are two suggestions, but you can come up with your own options as well:

• Try using several sizes of the set of documents and see how the time of building the structure and its memory footprint changes. (You can use subsets or multiple copies of the existing documents.)
• Try comparing queries with frequent words and those with less frequent words. How does the size of the output or intermediate lists in the multiquery influence the running time?
• Do not overload vyuka server too much with your experiments, also delete large input and output files when they are no longer needed.
• Create some plots showing the obtained results. Submit figures with these plots (png or pdf format) and discuss them in the protocol.
• Also describe in the protocol what commands you used to create your input files. If you used some scripts for this purpose, submit them.