In order to test and compare the performance of programs compiled using the enhanced pseudo language framework a simple program with a number of computations within a test for loop is implemented using both the traditional FOR loop and also the PARALLELFOR loop. These programs were then organized and compiled with an increasing number of loop iterations (and data elements) to compare/analyze the execution time as a function of elements/(loop size). The CUDA access library mentioned above was compiled with a hard-coded thread block size of 100 specified in a linear block definition so the loops smaller than 100 load a number of threads greatly exceeding those needed. These tests have been compiled and executed using test sizes ranging from 0 to 100000 and the results are illustrated in the chart below.

The figure above graphically shows the execution time for both the serial (baseline) and parallel execution modes. From the chart, one can easily see that the serial execution time is linearly proportional to the number of data elements/iterations, where as the parallel execution are essentially constant. It is important to note here that the pseudo language has no support for outputting the system time so the time markers used are outside the execution of the java VM within the test script itself. Because of this, the java loading overhead is included in the execution times for both the serial and parallel execution results. Also, the parallel versions incur a one-time overhead for loading the shared libraries. This overhead should not be too significant, but it should be noted in the results. In any event, the parallel execution mode demonstrates good performance compared to t he serial versions. Considering the lack of optimizations and a somewhat inefficient data movement design this is quite remarkable. Had the purpose of this effort been to optimize the performance results, these figures could have been drastically improved.