Amber 10 Benchmarks

By Ross Walker (SDSC)

| JAC Benchmark (PME, 23.5K atm) | Factor IX Benchmark (PME, 91K atm) | Cellulose Benchmark (PME, 408K atm) |
| Download Benchmark Suite Input Files |

Notes:

• You should not take any set of timings all that seriously. The values below are mainly useful to see generally how one machines performs against another, or how much (or little) parallel speedup we obtain for these systems. You should expect to get roughly similar performance on your Amber 10 installation with similar machines.
   
• We welcome additions or suggestions, and will continue to update this page from time to time.
   
• The timings on this page are for PMEMD v10.0 unless otherwise noted.
   
• Parallel speedups are generally highly dependent on the processor speed, and on the hardware and software details of the communications system. Note, parallel scaling normally improves as the system size increases.
   
• Nanoseconds per day is 86400/(seconds per nanosecond).
   
• When running on multi-core systems it can sometimes be beneficial to leave some of the cores idle (even though the supercomputer centers charge you for all of the cores). E.g. on NICS Kraken the best performance is achieved when only 1 MPI thread per node is run (leaving 3 or the cores idle). The difference in performance as a function of the number of cores per node (PPN) used is shown in the tables below.
   
• These benchmarks are part of a benchmark suite for AMBER prepared by Ross Walker. This suite contains the default JAC, FactorIX and Cellulose benchmarks but additionally contains more realistic versions of these benchmarks that reflect options that most people would use when running production MD. Also included are a series of implicit solvent Generalized Born benchmarks. A tar file containing the benchmark inputs can be downloaded here.
   
  • JAC - This is the original DHFR Joint Amber/Charmm Benchmark.
    23,558 atoms, explicit TIP3P solvent PME, 1fs time step, 9 A direct space cut off, NFFT=64x64x64, shake, shake_tol = 10^-7, NVE, ntpr = 100, ntwx = 0, nstlim = 1000.
     
  • JAC_production_NVE
    As JAC but a longer run with 2fs time step, NVE, 8 A direct space cut off, printing to mdout every 2ps, binary netcdf output to mdcrd every 2ps and dsum_tol and tol values appropriate for 'good' NVE energy conservation.
     
  • JAC_production_NVT/NPT
    As with JAC_production_NVE but with a 'very weak' Berendsen thermo or barostat respectively and with default dsum_tol and tol values.
     
  • FactorIX - This is the original FactorIX benchmark.
    90,906 atoms, explicit TIP3P solvent PME, 1.5fs time step, 8 A direct space cut off, shake, shake_tol = 10^-6, NVE, ntpr = 100, ntwx = 0, nstlim = 500
     
  • FactorIX_production_NVE
    As FactorIX but a longer run with 2fs time step, NVE, printing to mdout every 2ps, binary netcdf output to mdcrd every 2ps and dsum_tol and tol values appropriate for 'good' NVE energy conservation.
     
  • FactorIX_production_NVT/NPT
    As with FactorIX_production_NVE but with a 'very weak' Berendsen thermo or barostat respectively and with default dsum_tol and tol values.
     
  • Cellulose_production_NVE
    408609 atoms,  explicit TIP3P solvent PME, 2fs time step, 8 A direct space cut off, shake, shake_tol = 10^-6, dsum_tol = 10^-6, NVE, ntpr = 1000, ntwx = 1000, ioutfm = 1, nstlim = 10000.
     
  • Cellulose_production_NVT/NPT
    As with Cellulose_production_NVE but with a 'very weak' Berendsen thermo or barostat respectively and with default dsum_tol and tol values.

Machine Description

1. Dual x Quad Workstation - 2 x Intel E5462 CPU at 2.80 GHz, 1600MHz FSB, 24MB Cache, 4 x 2GB PC2-6400 FB ECC DIMMS, SuperMicro X7DWA-N Motherboard, RHEL4 AS
   Intel Ifort v10.1.018, MKL v10.0.5.025, mpich2-1.0.7 - PMEMD config.h.
    
2. Dual x Quad Workstation - 2 x Intel E5430 CPU at 2.66 GHz, 1333MHz FSB, 24MB Cache, 4 x 2GB PC2-5300 FB ECC DIMMS, SuperMicro X7DWA-N Motherboard, RHEL4 AS
   Intel Ifort v10.1.018, MKL v10.0.5.025, mpich2-1.0.7 - PMEMD config.h.
    
3. TACC Ranger - NSF Supercomputer at University of Texas at Austin, Sun Constellation Cluster, 579 TFlops Peak, 3,936 Nodes, 4 x Quad Core AMD Opteron Barcelona at 2.3 GHz, CentOS 4.7
   SDR Infiniband interconnect.
   Default environment: PGF90 v7.2-5, mvapich v1.0.1, tacc_affinity, ibrun - PMEMD config.h
    
4. NICS Kraken - NSF Supercomputer at University of Tennessee - Cray XT4, Quad core AMD nodes at 2.3GHz, Cray Compute Node Linux.
   Cray CStar2 Interconnect
   Default environment: ftn (PGF90 v7.1.6), xt-asyncpe v1.0.- PMEMD config.h
    

Joint Amber/Charmm DHFR Benchmark (JAC)

1) JAC Original Benchmark (Note: This does NOT do any I/O and so is not very representative of real calculations)
 

2) JAC Production NVE Benchmark (This is a more representative set of benchmarks for real world calculations)

3) JAC Production NVT Benchmark (This is a more representative set of benchmarks for real world calculations)

4) JAC Production NPT Benchmark (This is a more representative set of benchmarks for real world calculations)

Factor IX Benchmark

1) FactorIXOriginal Benchmark (Note: This does NOT do any I/O and so is not very representative of real calculations)

2) FactorIX Production NVE Benchmark (This is a more representative set of benchmarks for real world calculations)

3) FactorIX Production NVT Benchmark (This is a more representative set of benchmarks for real world calculations)

4) FactorIX Production NPT Benchmark (This is a more representative set of benchmarks for real world calculations)

Cellulose Benchmark

1) Cellulose Production NVE Benchmark (This is a more representative set of benchmarks for real world calculations)

2) Cellulose Production NVT Benchmark (This is a more representative set of benchmarks for real world calculations)

3) Cellulose Production NPT Benchmark (This is a more representative set of benchmarks for real world calculations)