TACC and NAMD

NAMD is currently installed on TACC's Frontera, Stampede2, and Lonestar6 compute resources. NAMD is managed under the module system on TACC resources. Read the following instructions carefully. NAMD performance is particularly sensitive to its configuration. Try running benchmarks with different configurations to find your optimal NAMD set up. You can initiate interactive idev debugging sessions on all systems.

Running on Frontera

The recommended and latest installed NAMD version is 2.14 on Frontera. Users are welcome to install different NAMD versions in their own directories.

login1$ module load namd/2.14

Job Scripts: NAMD on Frontera

TACC staff recommends that users attempt runs with 4 tasks per node and 8 tasks per node (scales better at large number of nodes) and then pick the configuration that provides the best performance.

4 tasks per node:

#SBATCH -J test         # Job Name
#SBATCH -o test.o%j
#SBATCH -N 2            # Total number of nodes
#SBATCH -n 8            # Total number of mpi tasks
#SBATCH -p normal       # Queue (partition) name 
#SBATCH -t 24:00:00     # Run time (hh:mm:ss) - 24 hours

module load namd/2.14
ibrun namd2 +ppn 13 \
            +pemap 2-26:2,30-54:2,3-27:2,31-55:2 \
            +commap 0,28,1,29 input &> output

8 tasks per node:

#SBATCH -J test         # Job Name
#SBATCH -o test.o%j
#SBATCH -N 12           # Total number of nodes
#SBATCH -n 96           # Total number of mpi tasks
#SBATCH -p normal       # Queue (partition) name -- skx-normal, skx-dev, etc.
#SBATCH -t 24:00:00     # Run time (hh:mm:ss) - 24 hours

module load namd/2.14
ibrun namd2 +ppn 6 \
            +pemap 2-12:2,16-26:2,30-40:2,44-54:2,3-13:2,17-27:2,31-41:2,45-55:2\
            +commap 0,14,28,42,1,15,29,43 input &> output

For very large simulations, users may want to use compressed structures. See the NAMD wiki: NamdMemoryReduction to prepare your compressed input files and set up your input files. In this case use the "namd2_memopt" executable instead of "namd2":

compressed input files

ibrun namd2_memopt +ppn 6 \
                +pemap 2-12:2,16-26:2,30-40:2,44-54:2,3-13:2,17-27:2,31-41:2,45-55:2 \
                +commap 0,14,28,42,1,15,29,43 input &> output

Running on Stampede2

As of this date, the recommended and latest version is 2.14 . Users are welcome to install different NAMD versions in their own directories. See Building Third Party Software in the Stampede2 User Guide.

Job Script: NAMD on Stampede2's KNL Nodes

TACC staff recommends assigning 13 tasks per node for NAMD jobs running on Stampede2's KNL compute nodes.

This job script requests 1 node and 4 MPI tasks: 4 tasks/node.

#SBATCH -J test         # Job Name
#SBATCH -o test.o%j
#SBATCH -N 1            # Request 1 node
#SBATCH -n 4            # and 4 MPI tasks
#SBATCH -p normal       # Queue (partition) name -- normal, development, etc.
#SBATCH -t 24:00:00     # Run time (hh:mm:ss) - 24 hours

module load namd/2.14
ibrun namd2_knl +ppn 32 \
                +pemap 0-63+68 \
                +commap 64-67 input &> output

To run the same job on more than one node, vary the "-N" and "n" #SBATCH directives. This job script requests 3 nodes and 39 MPI tasks: 13 tasks/node.

#SBATCH -J mynamd       # Set job name
#SBATCH -o mynamd.o%j
#SBATCH -N 3            # Request 3 nodes
#SBATCH -n 39           # Request 39 MPI tasks; 13 tasks per node
#SBATCH -p normal       # Submit to the normal queue
#SBATCH -t 24:00:00     # Set max run time of 24 hours

module load namd/2.14
ibrun namd2_knl +ppn 8 \
                +pemap 0-51+68 \
                +commap 52-67 input &> output

As well as the Slurm #SBATCH directives ("-N" and "-n"), try varying the affinity settings to determine the optimal performance of your job. You can try both settings then use the optimal one. If your system is small or the number of nodes are large, you can try:

4 tasks per node:

ibrun namd2_knl +ppn 16 +pemap 0-63 +commap 64-67

13 tasks per node:

ibrun namd2_knl +ppn 4 +pemap 0-51 +commap 52-67

Job Script: NAMD on Stampede2's SKX Nodes

TACC staff recommends assigning 4 tasks per node for jobs running on Stampede2's SKX compute nodes.

#SBATCH -J test         # Job Name
#SBATCH -o test.o%j
#SBATCH -N 2            # Total number of nodes
#SBATCH -n 8            # Total number of mpi tasks
#SBATCH -p skx-normal   # Queue (partition) name -- skx-normal, skx-dev, etc.
#SBATCH -t 24:00:00     # Run time (hh:mm:ss) - 24 hours

module load namd/2.14
ibrun namd2_skx +ppn 11 \
                +pemap 2-22:2,26-46:2,3-23:2,27-47:2 \
                +commap 0,24,1,25 input &> output

You may also try other affinity settings as in these examples for varying number of tasks per node.

6 tasks per node:

ibrun namd2_skx +ppn 7 \
                +pemap 2-14:2,18-30:2,34-46:2,3-15:2,19-31:2,35-47:2 \
                +commap 0,16,32,1,17,33 input &> output  

2 tasks per node:

ibrun namd2_skx +ppn 23 \
                +pemap 2-47:2,3-47:2 \
                +commap 0,1 input &> output

1 task per node:

ibrun namd2_skx +ppn 47 \
                +pemap 2-47:2,1-47:2 \
                +commap 0 input &> output

Running NAMD on Lonestar6

NAMD ver2.14 is installed on Lonestar6 as this version provides best performance. Feel free to install your own newer version locally.

login1$ module load namd/2.14

Job Script: NAMD on Lonestar6

TACC staff recommends assigning 4 tasks per node for NAMD jobs running on Lonestar6's compute nodes.

The following Lonestar6 job script requests 2 node and 8 MPI tasks. To run the same job on more nodes, vary the "-N" and "-n" Slurm directives, ensuring the value of n is four times the value of N.

#!/bin/bash
#SBATCH -J test         # Job Name
#SBATCH -o test.o%j
#SBATCH -N 2            # Total number of nodes
#SBATCH -n 8            # Total number of mpi tasks
#SBATCH -p normal       # Queue name
#SBATCH -t 24:00:00     # Run time (hh:mm:ss) - 24 hours

module load namd/2.14
ibrun namd2 +ppn 31 \
            +pemap 1-31,33-63,65-95,97-127 \
            +commap 0,32,64,96 input &> output

References

• NAMD website
• NAMD 2.14 User Guide
• TACC Software page