how to speed up this code? calculus iterating through rows and cols of a matrix

Question

I have a piece of code which performs calculations on a matrix by iterating through its rows and columns. The piece of calculus performed is a cosine distance measure, with a code I found on Internet (could not retrieve the link right now).

There can be 10,000 rows and col. The matrix is symmetric so I just need to iterate half of it. Values are float.

The problem: it is very slow (will take 3 to 6 hours it seems). Can anyone point me to improvements? Thx!

Note on the code: it uses an abstract class for flexibility: this way, the cosine calculation defined in a separate class could be easily replaced by another one.

The code:

import Jama.Matrix;
import java.util.ArrayList;
import java.util.HashSet;
import java.util.concurrent.ExecutionException;

public abstract class AbstractSimilarity {

    HashSet<Triple<Double, Integer, Integer>> set = new HashSet();
    public ArrayList<Thread> listThreads = new ArrayList();

    public void transform(Matrix matrixToBeTransformed) throws InterruptedException, 
ExecutionException {

        int numDocs = termDocumentMatrix.getColumnDimension();

        Main.similarityMatrix = new Matrix(numDocs, numDocs);

        System.out.println("size of the matrix: " + numDocs + "x " + numDocs);

        //1. iteration through all rows of the matrixToBeTransformed
        for (int i = numDocs - 1; i >0 ; i--) {
            System.out.println("matrix treatment... " + ((float) i / (float) numDocs * 100) + "%");

            //2. isolates the row i of this matrixToBeTransformed
            Matrix sourceDocMatrix = matrixToBeTransformed.getMatrix(
                    0, matrixToBeTransformed.getRowDimension() - 1, i, i);



            // 3. Iterates through all columns of the matrixToBeTransformed
//            for (int j = 0; j < numDocs; j++) {
//                if (j < i) {
//
//                    //4. isolates the column j of this matrixToBeTransformed
//                    Matrix targetDocMatrix = matrixToBeTransformed.getMatrix(
//                            0, matrixToBeTransformed.getRowDimension() - 1, j, j);


                    //5. computes the similarity between this given row and this given column and writes it in a resultMatrix
//                    Main.resultMatrix.set(i, j, computeSimilarity(sourceDocMatrix, targetDocMatrix));
//                } else {
//                    Main.resultMatrix.set(i, j, 0);

//                }
//
//            }
        }

The class which defines the computation to be done:

import Jama.Matrix;

public class CosineSimilarity extends AbstractSimilarity{

  @Override
  protected double computeSimilarity(Matrix sourceDoc, Matrix targetDoc) {
    double dotProduct = sourceDoc.arrayTimes(targetDoc).norm1();
    double eucledianDist = sourceDoc.normF() * targetDoc.normF();
    return dotProduct / eucledianDist;
  }

}

Is this a homework project? Can you not use math software such as MatLab? — Dan W, Mar 02 '12 at 15:20
This is a professional project in academia and I need to use Java for it - bc of my own limitations I'm afraid! — seinecle, Mar 02 '12 at 15:24
Have you profiled to see what part of your algorithm takes the longest time? Simply adding `new Date().getTime();` in the beginning/end of operations and them subtracting them can give you a nice insight. — Marcelo, Mar 02 '12 at 15:25
I also recommend using JVisualVM to see if its memory bound and the GC has to do a lot of work. Then you could for example allocate the doubles in your CosineSimilarity class up front, to see if that reduces memory fluctuation. — devsnd, Mar 02 '12 at 16:08

score 2 · Accepted Answer · answered Mar 02 '12 at 16:17

You appear to be dealing with a n^3 algorithm. n^2 because you're filling a (half) matrix. Times n again because the methods to fill each element(dot product/fnorm) take time n. The good news is that because the calculations don't depend on each other, you could multithread this to speed it up.

public class DoCalc extends Thread
{
  public Matrix localM;
  int startRow;
  int endRow;
  public DoCalc(Matrix mArg, int startArg, int endArg)
  {
    localM=mArg;
    startRow=startArg;
    endRow=endArg;
  }

  public void doCalc()
  {
    //Pseudo-code
    for each row from startRow to endRow
      for each column 0 to size-1
        result[i][j] = similarityCalculation
  }
  public void run()
  {
    doCalc();
  }
}

public void transform(Matrix toBeTransformed)
{
  int numDocs = termDocumentMatrix.getColumnDimension();

  Main.similarityMatrix = new Matrix(numDocs, numDocs);
  Vector<DoCalc> running = new Vector<DoCalc>();
  int blockSize = 10;
  for (int x = 0; x < numDocs-1;x+=blockSize)
  {
    DoCalc tempThread = new DoCalc(toBeTransformed,x,(x+blockSize>numDocs-1)?numDocs-1:x+blockSize);
    tempThread.start();
    running.add(tempThread);
  }

  for (DoCalc dc : running)
    dc.join();

}

Important notes:

This is a very naive implementation. If you try to run it with arrays of your size, it will spawn 1000 threads. You can either fiddle with the blockSize or look into thread pooling.

At best this will give you a multiple times increase in speed, 4x etc. If you want order of magnitude gains, you will need to properly profile and/or change your algorithm to something more efficient. Given the task you're trying to perform(running a relatively expensive task on each element in a Matrix), the latter may not be possible.

Edit: Multithreading will only increase speed significantly if you are cpu bound and have a multicore cpu with cores sitting relatively idle.

thx! a multithreading solution with a fixed pool of threads speed up my code by a factor 3. It still took 3h to complete on a 9400x9400 matrix. Now, I am investigating solutions to get way more speed! :-) => http://stackoverflow.com/questions/9550486/tutorials-or-books-on-kernel-programming-for-opencl — seinecle, Mar 04 '12 at 21:30

how to speed up this code? calculus iterating through rows and cols of a matrix

1 Answers1