Edit

Deploy NAMD on a virtual machine

Azure Virtual Machines
Azure Virtual Network

Caution

This article references CentOS, a Linux distribution that is nearing End Of Life (EOL) status. Please consider your use and plan accordingly. For more information, see the CentOS End Of Life guidance.

This article describes the steps for running NAMD software on a virtual machine (VM) that's deployed on Azure. It also presents the performance results of running NAMD on single-node and multi-node VM configurations.

NAMD is a computer application for molecular dynamics simulation that's based on the Charm++ parallel programming model. It's often used to simulate systems that comprise millions of atoms. NAMD supports hundreds of cores for typical simulations and can support more than 500,000 cores for the largest simulations. Simulations and trajectory analysis are performed with the popular molecular graphics program VMD, but NAMD is also compatible with AMBER, CHARMM, and X-PLOR. A source code version of NAMD is available for free.

NAMD is used mainly for high-performance simulations of large biomolecular systems. Typical NAMD simulations include all-atom models of proteins, lipids, nucleic acids, and explicit solvents (water and ions).

Why deploy NAMD on Azure?

  • By running NAMD on Azure HB-series VMs, you can reduce the time and cost of your simulations.
  • Running molecular simulation and analysis tasks on Azure can make it easier to implement advanced simulation methods and practical solutions for many molecular modeling tasks.
  • NAMD and associated tools enable popular research workflows like MDFF structure refinement and QwikMD simulation protocols to be run remotely. You don't need to invest in local computing resources, and the required expertise in high-performance computing (HPC) technologies is reduced.

Architecture

This diagram shows a multi-node configuration:

Diagram that shows a multi-node configuration.

Download a Visio file of this architecture.

This diagram shows a single-node configuration:

Diagram that shows a single-node configuration.

Download a Visio file of this architecture.

Components

Compute sizing and drivers

HBv3 AMD EPYC 7V73X (Milan-X) VMs running Linux CentOS were used to test the performance of NAMD on Azure. The following table provides details about HBv3-series VMs.

Size vCPU Memory (GiB) Memory bandwidth (GBps) Base CPU frequency (GHz) All-cores frequency (GHz, peak) Single-core frequency (GHz, peak) RDMA performance (GBps) Maximum data disks
Standard_HB120rs_v3 120 448 350 1.9 3.0 3.5 200 32
Standard_HB120-96rs_v3 96 448 350 1.9 3.0 3.5 200 32
Standard_HB120-64rs_v3 64 448 350 1.9 3.0 3.5 200 32
Standard_HB120-32rs_v3 32 448 350 1.9 3.0 3.5 200 32
Standard_HB120-16rs_v3 16 448 350 1.9 3.0 3.5 200 32

Required drivers

To use InfiniBand, you need to enable InfiniBand drivers.

Install NAMD 2.14 on a VM or HPC cluster

You can download the software from the NAMD website. You just need to untar or unzip the NAMD binary distribution file and run it in the resulting directory. For information about building from source code, see Compiling NAMD.

Before you install NAMD, you need to deploy and connect to a VM or an HPC cluster.

For information about deploying the VM and installing the drivers, see Run a Linux VM on Azure.

For information about deploying Azure CycleCloud and the HPC cluster, see these articles:

NAMD performance results

Two models were used to test the scalability performance of NAMD 2.14 on Azure:

  • STMV. A small icosahedral plant virus that worsens the symptoms of infection by tobacco mosaic virus (TMV).
  • ApoA1. A component of high-density lipoprotein (HDL). The ApoA1 gene provides instructions for creating a protein called apolipoprotein A-I.

The details about each test model are provided in the following sections.

Model 1: STMV

Screenshot that shows an STMV model.

To validate NAMD 2.14 scaling on Azure HPC systems, STMV is tiled in arrays of 5x2x2 (21 million atoms). The following table provides details about two versions of the model.

Model Number of atoms Time step Number of steps Method
1a 1,066,628 1 2,000 Particle-Mesh Ewald (PME)
1b 21,000,000 2 1,200 PME

Model 2: ApoA1

Screenshot that shows the ApoA1 model.

The following table provides details about the model.

Model Number of atoms Time step Number of steps Method
2 92,224 1 2,000 PME

NAMD 2.14 performance results on single-node VMs

The following sections provide the performance results of running NAMD on single-node Azure HBv3 AMD EPYC 7V73X (Milan-X) VMs.

Model 1a: STMV

This table shows the nanoseconds per day and total wall-clock times recorded for varying numbers of CPUs on the Standard HBv3-series VM:

Number of cores ns per day Wall-clock time (seconds) Relative speed increase
16 0.168 1,046.34 NA
32 0.306 633.58 1.65
64 0.535 380.58 2.75
96 0.692 308.60 3.39
120 0.692 312.93 3.34

The following graph shows the relative speed increases as the number of CPUs increases:

Graph that shows the relative speed increases.

The following graph shows the nanoseconds-per-day for varying numbers of CPUs:

Graph that shows the nanoseconds-per-day for varying numbers of CPUs.

Model 1b: STMV

This table shows the nanoseconds per day and total wall-clock times recorded for varying numbers of CPUs on the Standard HBv3-series VM:

Number of cores ns per day Wall-clock time (seconds) Relative speed increase
16 0.014 14,712.03 NA
32 0.028 7,715.94 1.91
64 0.054 4,092.31 3.60
96 0.070 3,239.06 4.54
120 0.078 2,955.45 4.98

The following graph shows the relative speed increases as the number of CPUs increases:

Graph that shows the relative speed increases for model 1b.

The following graph shows the nanoseconds-per-day for varying numbers of CPUs:

Graph that shows the nanoseconds-per-day increase for model 1b.

Model 2: ApoA1

This table shows the nanoseconds per day and total wall-clock times recorded for varying numbers of CPUs on the Standard HBv3-series VM:

Number of cores ns per day Wall-clock time (seconds) Relative speed increase
16 1.794 130.61 NA
32 3.004 82.67 1.58
64 4.285 66.36 1.97
96 4.966 61.51 2.12
120 5.288 58.89 2.22

The following graph shows the relative speed increases as the number of CPUs increases:

Graph that shows the relative speed increases for the ApoA1 model.

The following graph shows the nanoseconds per day for varying numbers of CPUs:

Graph that shows the nanoseconds-per-day increase for the ApoA1 model.

Notes about the single-node tests

For all single-node tests, the solver time on a Standard_HB120-16rs_v3 VM (16 cores) is used as a reference to calculate the relative speed increase with respect to similar VMs that have more cores. The previously presented results show that parallel performance improves as cores increase from 16 to 64. At 120 cores, the improvement is limited and only occurs on some simulations. This pattern is common with these simulations and other memory-intensive applications because of the saturation of the onboard memory that's available on each processor. Taking VM costs into consideration, the 64-CPU configuration is the best choice. Standard_HB120-64rs_v3 VMs, which have 64 cores, were used for the multi-node tests.

NAMD 2.14 performance results on a multi-node cluster

The single-node tests confirm that the solver achieves optimal parallel performance with 64 cores on HBv3-series VMs. Based on those results, 64-core configurations on Standard_HB120-64rs_v3 VMs were used to evaluate the performance of NAMD on multi-node clusters. STMV model 1b was used for the multi-node tests.

This table shows the nanoseconds per day and total wall-clock times recorded for varying numbers of nodes on Standard HBv3-series VMs:

Number of nodes Number of cores ns per day Wall-clock time (seconds) Relative speed increase
1 64 0.054 3,835.48 NA
2 128 0.157 1,340.54 2.86
4 256 0.313 675.18 5.68
8 512 0.619 346.97 11.05
16 1,024 1.221 183.09 20.95

The following graph shows the relative speed increases as the number of nodes increases:

Graph that shows the relative speed increases on multi-node clusters.

The following graph shows the nanoseconds per day for varying numbers of nodes:

Graph that shows the nanoseconds per day on multi-node clusters.

The results show that model 1b scales well as the number of nodes increases. For better performance, use a memory-optimized version of the application. You can optimize the application by compiling the source code. The simulation used for testing is limited to a few iterations. Because real-world applications can use more iterations, you can minimize the total time that's required for decomposition, which further improves performance.

Azure cost

The following table provides wall-clock times that you can use to calculate Azure costs. To compute the cost, multiply the solver running time by the number of nodes and the Azure VM hourly cost. For the current hourly costs, see Linux Virtual Machines Pricing. The Azure VM hourly rates are subject to change.

Only simulation running time is considered for the cost calculations. Installation time, simulation setup time, and software costs aren't included.

You can use the Azure pricing calculator to estimate VM costs for your configurations.

STMV model 1b was used to calculate these times.

Number of nodes Wall-clock time (hours)
1 1.07
2 0.37
4 0.19
8 0.10
16 0.05

Summary

  • NAMD 2.14 was successfully tested on Azure HBv3 standalone VMs and on an Azure CycleCloud multi-node configuration.
  • Model 1b scales well on the multi-node configuration. On a 16-node configuration, the speed is 21 times faster than it is on a single node.
  • For better performance, we recommend that you use the +p option to run one thread per processor. We also recommend that you look for prebuilt ibverbs NAMD binaries or specify ibverbs when you build the Charm++.
  • For small simulations, we recommend that you use fewer CPUs.

Contributors

This article is maintained by Microsoft. It was originally written by the following contributors.

Principal authors:

Other contributors:

To see non-public LinkedIn profiles, sign in to LinkedIn.

Next steps