Namd on BGQ
A parameter study was undertaken to test simulation performance and efficiency of NAMD on the Blue Gene/Q cluster, BGQ, with attention to NAMD performance tuning documentation. Determining optimal parameters for a NAMD simulation on this system is more difficult as there are only certain simulation sizes that have optimal topologies (512, 1024, etc). The system of study is a 246,000 atom membrane protein simulation (Cytochrome c Oxidase embedded in a TIP3P solvated DPPC bilayer) using the CHARMM36 forcefield (protein and lipids). The unit cell is cubic with box dimensions 144 x 144 x 117 Angstroms and the simulation time-step was 2fs. The following non-bonded frequency parameters were also used: nonbondedFreq=1, fullElectFrequency=2, stepspercycle=10.
The following benchmarks were performed with the non-SMP NAMD build with exception to the last section. All simulations are started using a restart file from a pre-equilibrated snapshot. Performance in nanoseconds per day is based on the geometric mean of the three "Benchmark time" lines at the beginning of the simulation's standard output and may not represent long-time averages.
Performance Tuning Benchmarks
Efficiency is measured with respect to the 16 ranks-per-node 512 core simulation. In this section, the PME patch grid was manually doubled in either the X, Y, or Z directions. Default PME patch doubling in NAMD 2.9 is generally recommended (twoAway parameters need not be specified in the configuration file).
| Ranks | Cores | NAMD Config Options | ns/day | Efficiency |
| 16 | 512 | 2.79 | 1.00 | |
| 16 | 1024 | 5.05 | 0.91 | |
| 16 | 1024 | twoAwayX (default) | 5.62 | 1.01 |
| 16 | 2048 | twoAwayX (default) | 10.07 | 0.90 |
| 16 | 2048 | twoAwayXY | 10.59 | 0.95 |
| 16 | 4096 | twoAwayX | 14.32 | 0.64 |
| 16 | 4096 | twoAwayXY (default) | 17.63 | 0.79 |
| 16 | 4096 | twoAwayXYZ | 16.79 | 0.75 |
| 16 | 8192 | twoAwayX | 23.52 | 0.53 |
| 16 | 8192 | twoAwayXY (default) | 25.00 | 0.56 |
| 16 | 16384 | twoAwayX | 23.67 | 0.27 |
| 16 | 16384 | twoAwayXY | 28.31 | 0.32 |
| 16 | 16384 | twoAwayXYZ (default) | 27.98 | 0.31 |
PME Pencils
A "pencil-based" PME decomposition may be more efficient than the default "slab-based decomposition". In this study PME pencil grids are created for both dedicated PME nodes (lblUnload=yes) and non-dedicated PME nodes. Fine-tuning of PMEPencils resulted in insignificant performance gains for this study.
| Ranks | Cores | NAMD Config Options | ns/day | Efficiency |
| 16 | 4096 | twoAwayXY, PMEPencils=8, lblUnload=yes | 12.93 | 0.58 |
| 16 | 4096 | twoAwayXY, PMEPencils=12, lblUnload=yes | 17.27 | 0.77 |
| 16 | 4096 | twoAwayXY, PMEPencils=16, lblUnload=yes | 16.02 | 0.72 |
| 16 | 4096 | twoAwayXY, PMEPencils=20, lblUnload=yes | 15.41 | 0.69 |
| 16 | 4096 | twoAwayXY, PMEPencils=12 | 16.21 | 0.73 |
| 16 | 4096 | twoAwayXY, PMEPencils=16 | 17.92 | 0.80 |
| 16 | 4096 | twoAwayXY, PMEPencils=20 | 17.99 | 0.81 |
| 16 | 4096 | twoAwayXY, PMEPencils=24 | 17.83 | 0.80 |
| 16 | 4096 | twoAwayXY, PMEPencils=36 | 16.97 | 0.76 |
| 8 | 4096 | twoAwayXY, PMEPencils=20 | 18.24 | 0.82 |
| 16 | 4096 | twoAwayXY, PMEPencils=20 | 17.99 | 0.81 |
| 32 | 4096 | twoAwayXY, PMEPencils=20 | 13.94 | 0.63 |
Ranks-Per-Node Study
The "ranks-per-node" or simply the number of processes per compute node is a Blue Gene/Q runjob command parameter. In this study, memory requirements were too large to use 64 due to memory errors, and also resulted in out of memory errors for 16384 core simulations of 32 ranks per node. The following efficiency estimates are measured with respect to the 16 ranks per node results for the same number of nodes respectively (for the default twoAway choices and no PME Pencils). These simulations offered the best performance and were used in production simulations. Note that 1024, the first entry in the table, means that 512 physical cores were requested, but due to the double ranks-per-node, a total of 1024 virtual cores were used.
| Ranks | Cores | NAMD Config Options | ns/day | Efficiency |
| 32 | 1024 | 4.46 | 1.6 | |
| 32 | 2048 | 8.01 | 1.43 | |
| 32 | 4096 | 13.74 | 1.3 | |
| 32 | 8192 | 19.81 | 1.12 |
Incorrect Particle-Mesh Ewald Grid
Long-range electrostatics are computed using PME for all simulations above with PME grid spacing set to be generated automatically with the "pmeGridSpacing 1.0" setting. A poor choice in PME grid spacing (i.e. not a multiple of 2,3, and 5) can result in increasingly large performance degradation due to the matrix size requirements in the FFT algorithm. Below is an example of the type of performance degradation that one may expect with none of the grid dimensions are divisible by 5. One can draw a comparison to a more correct PME choice in the Performance Tuning Benchmarks above.
| Ranks | Cores | NAMD Config Options | ns/day | Efficiency |
| 16 | 512 | Poor PME Multiple (144x144x111) | 2.70 | 0.97 |
| 16 | 1024 | Poor PME Multiple (144x144x111) | 5.13 | 0.92 |
| 16 | 2048 | Poor PME Multiple (144x144x111) | 8.61 | 0.77 |
| 16 | 4096 | Poor PME Multiple (144x144x111) | 13.93 | 0.62 |
| 16 | 8192 | Poor PME Multiple (144x144x111) | 17.08 | 0.38 |
| 16 | 16384 | Poor PME Multiple (144x144x111) | 17.64 | 0.20 |
Symmetric multiprocessing (SMP) Study
All the prior benchmarks were performed with a non-SMP NAMD binary, thus true multithreading was not possible. The following results utilize this feature and offer performance increases of up to 40% compared to the non-SMP benchmarks. Processors per node (ppn) was varied from 1 (not multithreaded) to 4. Values greater than 4 resulted in a crash. The addition of the Charm++ flag "+CmiNoProcForComThread" to the NAMD binary had negligible improvements in all cases. Comparing 512/ppn4 with 1024/ppn4 we observed 85% scaling, within acceptable criteria for production runs. Efficiency is respect to the 512/ppn1 simulation.
| Ranks | Cores | NAMD Config Options | ns/day | Efficiency |
| 16 | 512 | ppn 1 | 2.75 | 1.00 |
| 16 | 512 | ppn 2 | 4.60 | 1.67 |
| 16 | 512 | ppn 3 | 5.59 | 2.03 |
| 16 | 512 | ppn 4 | 6.31 | 2.29 |
| 16 | 512 | ppn 1, +CmiNoProcForComThread | 2.74 | 1.00 |
| 16 | 512 | ppn 2, +CmiNoProcForComThread | 4.62 | 1.68 |
| 16 | 512 | ppn 3, +CmiNoProcForComThread | 5.50 | 2.00 |
| 16 | 512 | ppn 4, +CmiNoProcForComThread | 6.31 | 2.30 |
| 16 | 1024 | ppn 1 | 5.54 | 1.01 |
| 16 | 1024 | ppn 2 | 8.31 | 1.51 |
| 16 | 1024 | ppn 3 | 8.56 | 1.56 |
| 16 | 1024 | ppn 4 | 10.60 | 1.92 |
| 16 | 1024 | ppn 4, +CmiNoProcForComThread | 10.58 | 1.92 |
| 16 | 2048 | ppn 1 | 9.93 | 0.90 |
| 16 | 2048 | ppn 2 | 13.86 | 1.26 |
| 16 | 2048 | ppn 3 | 14.72 | 1.34 |
| 16 | 2048 | ppn 4 | 15.10 | 1.37 |
The combination of increased ranks-per-node and mulithreading was not as effective as using the normal 16 ranks-per-node and ppn 4 as above. Simulations larger than ppn 2 were not successful.
| Ranks | Cores | NAMD Config Options | ns/day | Efficiency |
| 32 | 512 | ppn 1, +CmiNoProcForComThread | 4.45 | 1.62 |
| 32 | 512 | ppn 2, +CmiNoProcForComThread | 6.26 | 2.27 |
PME pencils offered minimal improvements when selected appropriately.
| Ranks | Cores | NAMD Config Options | ns/day | Efficiency |
| 16 | 512 | ppn 4, pmePencils 4 | 2.88 | 1.05 |
| 16 | 512 | ppn 4, pmePencils 16 | 6.34 | 2.31 |
| 16 | 512 | ppn 4, pmePencils 20 | 6.35 | 2.31 |
| 16 | 512 | ppn 4, pmePencils 24 | 6.24 | 2.27 |