Benchmarks

From CYANA Wiki
Revision as of 14:47, 10 February 2010 by Guentert (talk | contribs) (Undo revision 7755 by Guentert (Talk))
Jump to navigation Jump to search

Benchmark results for CYANA 3.0:

Computer system Number of processors (cores) Computation time (s) Relative performance
Basic Auto
Laptop 1 813 ? 0.26
Desktop (intel) 1 208 ? = 1
4 57 2479 3.65
Desktop (gfortran) 1 258 ? 0.81
Desktop (g95) 1 336 ? 0.62
Mac mini (intel) 1 ? ? ?
2 87 4092 2.39
Mac mini (gfortran) 2 104 4894 2.00
Server 1 146 ? 1.42
10 16 745 13.0
20 11 412 18.9
50 4 209 52.0
100 5 142 41.6

February 2010

Computer systems:
  • Laptop: Laptop computer with 1 Intel Core 2 Duo L7800 dual-core CPU, 2.00 Ghz, 2 GB memory, Cygwin under Windows 7, gfortran (-O3) Fortran compiler 4.3.4.
  • Desktop: Desktop computer with 1 Intel Core 2 Q9400 quad-core CPU, 2.66 Ghz, 4 GB memory, Ubuntu 8.10 Linux, Intel Fortran compiler 11.1, gfortran (-O3) Fortran compiler 4.3.2, or g95 Fortran compiler 0.91.
  • Mac mini: Mac mini with 1 Intel Core 2 Duo dual-core CPU, 2.53 Ghz, 4 GB memory, Mac OS X 10.6.2, Intel Fortran compiler 11.1 or gfortran (-O3) Fortran compiler 4.5.0.
  • Server: Linux cluster system with 20 nodes, each having 2 Intel Xeon E5462 quad-core CPUs, 2.80 Ghz, 16 GB memory, Ubuntu 8.04 Linux, Intel Fortran compiler 10.1, OpenMPI.
Benchmark types:
  • Basic: Structure calculation of a protein with 114 amino acid residues, 1737 NOE distance restraints, 110 torsion angle restraints. 50 conformers are calculated using 4000 torsion angle dynamics steps per conformer.
  • Auto: Structure calculation with automated NOESY assignment of a protein with 114 amino acid residues, 4732 NOESY cross peaks, 88 torsion angle restraints. 8 x 100 = 800 conformers are calculated using 10000 torsion angle dynamics steps per conformer.
Relative performance
  • Normalized to 1.0 for the calculation with the 'Desktop' computer system using 1 processor core.
  • Based on the computation time for the Basic calculation

Notes

  • Running the complete CYANA 3.0 test suite, including the 'basic' and 'auto' calculations, as well as structure calculations for homodimers, structure calculations using RDCs, pseudocontact shifts, etc., takes about 3.7 times as long as the Auto benchmark.