View source for Benchmarking NAMD workflows for GPU containers

<span id="benchmarking-namd2-workloads-for-gpu-containers-on-ccast"></span>
== Benchmarking NAMD2 workloads for GPU containers on CCAST ==

Stephen Szwiec 

Rasulev Computational Chemistry Research Group

North Dakota State University

26 July 2022
-----
==== ApoA1 information ====

* simulates a bloodstream lipoprotein particle
* '''physical stats'''
** 92224 atoms
** 70660 bonds
** 74136 angles
** 74130 diheadrals
** 1402 impropers
** 32992 hydrogen groups
** 553785 total amu of mass
* '''energy stats'''
** 300K initial temp.
** -14 e total charge
* '''simulation stats'''
** Consists of a startup process followed by 500 steps of simulation for benchmark time
** GPU workload modified to use CUDA FFTW, CUDA integration
** GPU workload modified to continue simulation for 10000 steps
==== NAMD non-gpu run information ====

* Charmrun used with one node (condo02)
* machine topology: 2sockets x 64 cores x 1 PU = 128-way SMP possible
* 20 processes, 20 cores, 1 physical node, and 16GB memory specified
* NAMD 2.14 used
==== NAMD gpu run information ====

* NAMD used within Singularity container system
** namd:3.0-alpha11 image used to generate singularity container
* machine topology: 2sockets x 64 cores x 1 PU = 128-way SMP possible
* 2 CPUs, 2 GPUs, 1 openMP thread, 1 physical node, and 16GB memory specified
** binding each GPU to one Nvidia A10 with 22731MB memory
* NAMD 3.0alpha11
-----
=== Benchmark findings ===
==== Startup Wall Time ====

'''CPU''' 3.4346 s '''GPU''' 0.0963 s
==== Simulation Wall Time ====

'''CPU''' 38.6991 s '''GPU''' 5.59571 s
==== Wall Time Per Step ====

'''CPU''' 0.0918308 s/step '''GPU''' 0.0110155 s/step
==== Days Per Nanosecond Simulation ====

'''CPU''' 0.999022 days/ns '''GPU''' 0.0660301 days/ns
-----
==== Additional information ====

* Extending Apoa1 workload steps with CUDA lead to step 3500 being reached in roughly the same time step 500 was reached on 20 CPUs.
* per core speedup with GPU acceleration is ~151.297 times
** caveat: one CPU must supervise and await output from each GPU because of how NAMD was written
* CUDA performed 10000steps and wrote output in 136.315155s total wall time
* CPU performed 500steps and wrote output in 76.477325s total wall time