444.namd
Jim Phillips <jim [at] ks.uiuc.edu>
Theoretical and Computational Biophysics Group
University of Illinois
Urbana-Champaign
Scientific, Structural Biology, Classical Molecular Dynamics Simulation
The 444.namd benchmark is derived from the data layout and inner loop of NAMD, a parallel program for the simulation of large biomolecular systems.
Although NAMD was a winner of a 2002 Gordon Bell award for parallel scalability, serial performance is equally important to the over 10,000 users who have downloaded the program over the past several years. Almost all of the runtime is spent calculating inter-atomic interactions in a small set of functions. This set was separated from the bulk of the code to form a compact benchmark for CPU2006. This computational core achieves good performance on a wide range of machines, but contains no platform-specific optimizations.
The file namd.input is a 92224 atom simulation of apolipoprotein A-I and is used as a standard NAMD benchmark. This particular file format is created by NAMD 2.5 using the "dumpbench" command, and eliminates the need for file readers and other setup code from the benchmark. Test, train and ref read from the same input file, but run the code for different number of iterations. For ref the code is run for 38 iterations.
CPU2006 uses the "--output namd.out" command-line option to produce an output file, namd.out, containing various checksums on the force calculations. These should be consistent across platforms, within roundoff error, and it is used for validation.
C++
The benchmark is written in conservative C++, is quite portable, and the inner loop code (module ComputeNonbondedUtil.C) contains no aliasing. The erfc() function is required for startup. On Windows, -DWIN32ERFC is defined during compilation to build a version of erfc() for little-endian, 32-bit and 64-bit platforms. This is only needed for startup, and should not affect overall performance.
Last updated: 27 June 2006