Server (paid or whatever) for running gromacs simulations

User discussions
1

Google doesn’t help…

Any recommendations for using paid or whatever server for running gromacs simulations?

Additionaly, as a solo researcher, new to the field, I’d like to know other’s experiences/solutions for running gromacs simulations:

  1. which server (paid or free for university members)
  2. how much nodes/cores you are able to use
  3. how much does it cost (if any)
  4. how long your typical simulations lasts…

I believe that even short answers on questions above would help newcomers like me…
Thanks

2

I can share my experiences as a researcher from the USA. Here in the USA, we get access to computational resources from followings,

  • Most of the Universities have their own supercomputing centers (HPC). Which are powerful enough to run 50-500K particle systems. I ran >500K particle system on our university’s HPC.

  • There are lots of national supercomputing centers, most of them have ranked in top500 list. These are very powerful, often you will find that million atoms simulation/ microsecond simulations are carried out using these resources. Generally, PIs write proposals to get computing time from these national HPCs.

  • Some research groups buy their own computing nodes. You will find, most of them are discrete computing nodes rather than distributed nodes. People skip building distributed computing nodes as, interconnectors (e.g. ethernet, Infiniband) are as expensive as computing nodes.

For independent researchers, GPU based solution is very viable and economic. Gromacs has an excellent performance track record on GPUs.

Let’s answer some of your points,

How much nodes/cores you are able to use?

There are no limits on how many cores you can use. It depends on your system size and type. You can use as many cores as possible till system’s computational performance scales. Performance of a code does not scale linearly, after a threshold core counts, you will see the plateau on Performance vs. cores plot, even degradation of performance. You should limit your core counts, such that it doesn’t hit a plateau on performance vs. core plot.

How long your typical simulations lasts?

It is a very generic question to answer. Again it depends on your system, and what question would you like to answer from the simulation. You need to check the convergence of your desired properties (e.g. Thermodynamic, structural, dynamic properties).

How much does it cost (if any)?

There are some cloud computing services that you can use, for example, AWS. I never used it, and I don’t think it is cheap and viable either. They charge based on nodes and hours. You might consider obtaining a GPU card as a viable/flexible alternative.

I hope you find information helpful.

Happy simulating,

Sincerely,
Masrul