Letter to the editor of Linux Journal

We were interested in reading the article by Michael Hirsch, "Comparing Java implementaions for Linux", p. 160 - 164, Linux Journal, August 2000. While he begins the "The Benchmark" section of the article with a disclaimer, about what this benchmark does not measure. He proceded to compare the Java benchmark results to a C++ implementation that looked similar. Unfortunately, the C++ version is quite different:

The benchmark is writen as if no GC is necessary. However, Java does garbage collection, while C++ does not. Thus the allocation of the PerformanceTester[NOBJECTS] array, and the insertion of each object into the array can trigger a GC at any time. The C++ version only allocates
The C++ version has no virtual methods to correspond to Java's instance methods.
The NOBJECTS parameter in C++ is defined as a constant, while in the Java version it is a nonconstant static field. This can add one or more instructions per loop.

This importance of each loop is quite different. For example, the fist loop is a consing loop, that accounts for about 80% of the time As show in the following exhibit:

3.8%

81.5%

11.5%

7.0%

public class PerformanceTester0 {
  private int amount;
  private static int NOBJECTS = 500000;
  public void addAmount(int val) { amount += val; }
  public int getAmount() { return amount; }

  public static void main(String[] args) {
    for (int i = 0; i < 9; i++)
      runTest(); }

  public static void runTest() {
    long startTime = 0;
    long endTime = 0;
    int sum = 0;

    startTime = System.currentTimeMillis();
    //  70 msec   3.8%
    PerformanceTester0[] ptarr = new PerformanceTester0[NOBJECTS]; 
    // 1493 msec 81.5%
    for (int i = 0; i < NOBJECTS; i++)
      ptarr[i] = new PerformanceTester0();
    //  210 msec  11.5%
    for (int i = 0; i < NOBJECTS; i++)
      ptarr[i].addAmount(i);
    //  129 msec   7.0%
    for (int i = 0; i < NOBJECTS; i++)
      sum += ptarr[i].getAmount();

    endTime = System.currentTimeMillis();
    System.out.println("Elapsed time in millisconds for "
		       + NOBJECTS + " objects; " + (endTime - startTime));
    System.out.println("Objects per millisecond: " +
		       NOBJECTS/(endTime - startTime));
  }
}

d:\silk4\silkweb\doc\Scheme2000>javac PerformanceTester0.java
javac PerformanceTester0.java

d:\silk4\silkweb\doc\Scheme2000>java PerformanceTester0
Elapsed time in millisconds for 500000 objects; 71
Elapsed time in millisconds for 500000 objects; 70
Elapsed time in millisconds for 500000 objects; 70
Elapsed time in millisconds for 500000 objects; 70 <-
Elapsed time in millisconds for 500000 objects; 70

d:\silk4\silkweb\doc\Scheme2000>javac PerformanceTester0.java
Elapsed time in millisconds for 500000 objects; 1702
Elapsed time in millisconds for 500000 objects; 1603
Elapsed time in millisconds for 500000 objects; 1582
Elapsed time in millisconds for 500000 objects; 1563 <-
Elapsed time in millisconds for 500000 objects; 1562

d:\silk4\silkweb\doc\Scheme2000>javac PerformanceTester0.java
Elapsed time in millisconds for 500000 objects; 1732
Elapsed time in millisconds for 500000 objects; 1732
Elapsed time in millisconds for 500000 objects; 1712
Elapsed time in millisconds for 500000 objects; 1703 <-
Elapsed time in millisconds for 500000 objects; 1702

d:\silk4\silkweb\doc\Scheme2000>javac PerformanceTester0.java
Elapsed time in millisconds for 500000 objects; 1853
Elapsed time in millisconds for 500000 objects; 1842
Elapsed time in millisconds for 500000 objects; 1843
Elapsed time in millisconds for 500000 objects; 1832 <-
Elapsed time in millisconds for 500000 objects; 1823
*/

Letter to the editor of Linux Journal

The Problem