Simple Run-time benchmarks for a Molecular Dynamics Simulation

Using OpenMp directives shown at:
http://www.openmp.org/index.cgi?samples+samples/md.html

This code was converted to Standard Fortran 90. All REAL
variables are Double Precision.

See below for MDYN.F90 Code Changes; changed version is available at:
http://ftp.cac.psu.edu/pub/ger/fortran/mdyn.f90
This report is available at:
http://ftp.cac.psu.edu/pub/ger/fortran/mdyn.txt


TEST RESULTS

Platform            Time        (Compiler)
------------------------------------------------
Dell P360            151 seconds(** Lahey)
Lion-XM:             155 seconds(* Lahey)
Lion-XL:             194 seconds(* Lahey)
Dell Dimension 8200   40 seconds(IFORT81)
AMD Opteron           95 seconds(PGI 6.0-2)
Hammer                11 seconds(IFORT64)
Hammer                 9 seconds(PathScale 2.1)
------------------------------------------------
seconds are from intrinsic function System_Clock.


Compilers and Options Used
--------------------------
(IFORT81) ifort 8.1.018 options were: ifort -O3 -xN -iop -static mdyn.f90
(IFORT64) ifort 64-bit version 9.0: ifort mdyn.f90 -O3 -xW -ipo -static -V -openmp
(* Lahey) Compiled on Lion-XL using: LF95 --o1 -static -openmp
   and ported to Lion-XM as a 32-bit application.
(** Lahey) Compiled on one of my Dell PC's using: LF95 -o1 -staticlib
    and ported to Dell Precision 360 in Room 214D.
PathScale 2.1 options:
 -Ofast -OPT:fast_math=on -WOPT:if_conv=off -LNO:fu=9:full_unroll_size=7000 -openmp
Portland 6.0-2 options:
 -Bstatic -V -fastsse -Munroll=n:4 -Mipa=fast,inline -tp k8-64


MDYN.F90 Code Changes:
----------------------
1) Use F90 Double Precision via Select Real Kind.
2) Add algorithm timing via System_Clock.
3) Replace call to compiler dependent call to random_number intrinsic with a call
   to own RAND subprogram so results can be duplicated across compilers.
4) Make v and dv external functions (else error in -openmp.
5) All reults then agreed to 5 decimal places.

The above times represent a simple average of THREE "runs"
The "runs" were local, that is NOT batch jobs.

Time is real time seconds from the system clock via Fortran
code at the beginning and end of the simulation.


Penn State GEaRS Group
29 June 2005