In February we released an update of the 3D AMaze programs for 64 bit multicore computers. Time-intensive activities are divided between multiple processors to give a substantial reduction in run time. The programs (compiled with Intel FORTRAN 11) incorporate OpenMP routines (http://openmp.org/wp/). A previous article described benchmark tests of the Magnum program (http://fieldp.com/myblog/2009/magnum-in-parallel/). We have created a test program to enable potential customers to check their machine performance. This article describes how to obtain and to use the program and summarizes benchmark tests we have carried out.
To perform the test, your computer must be running a 64-bit version of Windows XP, Vista or 7. Use this links to obtain the program:
http://www.fieldp.com/myblog/examples/openmptest.exe
Run openmptest to display the dialog shown below. The field Number of threads initially displays the maximum number of threads on your computer. Change the number to determine how many threads should be used in the test.
[caption id="" align="aligncenter" width="294"]
Screen display, openmptest.exe[/caption]
You will probably want to run the Windows Task Manager to display computer resource allocation during the test. Push CRTL-ALT-DELETE and select the performance tab. In openmptest, click the button Run test to check the machine speed with the specified number of threads. The activity (a large loop of computations with double precision numbers) is an ideal task for a multiprocessor machine. There are no I/O tasks and all operations are performed in parallel. In principle, the run time should be given by
TRun = T1/N
where T1 is the run time for serial operation and N is the number of parallel paths.
I obtained the following results on an older dual Xeon workstation (HP XW6200) running at 3.2 GHz under Windows XP. The time with a single processor was 25.42 s, while the time with both processors was 12.93 s. The reduction factor of 1.97 was close to the theoretical value. On a new computer with an Intel Core I7 processor at 1.60 GHz (8 thread, quad core), the run time using a single thread was 12.05 s. Despite the lower clock speed, the new processor was much faster. With all 8 threads in play, the run time dropped to 3.26 s. The figure below shows a plot of run time versus the number of threads. The dashed green line is the theoretical value. The maximum reduction factor was about 4.0 rather than 8.0. The run time reduction was determined by the effective number of processors rather than the number of individual threads.
CAPTION
LINKS
