What are some cloud services for computing?

Question

A group of researchers in my university (including me), started to study the possibility to move to cloud computing services in order to run our simulations.

The simulation areas include condensed matter physics, particle physics, quantum field theory, soft condensed matter, statistical physics, etc.

We have zero experience with it.

The question is: Are there reliable cloud computing services for matter modeling?

By reliable, I mean that their performance (CPU/GPU use, RAM use, hard drive speed) is similar to a local computer, with lower degradation (as many of them use virtual machines). As the demands are very heterogeneous, at first glance, we are thinking in bare iron system in order to install/configure different software for running the simulations. About the hardware, it is desirable to have both CPUs and GPUs access.

To distinguish this question from these ones:

• Supercomputers around the world
• Supercomputers around the world, which allow access for researchers outside the country,

where access is usually not just granted to anyone who pays for it, perhaps you can suggest commercial services here.

Answer 1

Materials Square (MatSQ)

Materials Square (MatSQ) is a web-based simulation platform providing an interface, workflow, and pay-as-you-go cloud server for material simulations. It supports DFT, MD and CALPHAD calculations.

• Provides a free 3D atomic structure-modeling interface supporting most structure types.

• Support density functional theory calculations of Electronic Structure, Phonon, reaction path and optical properties using Quantum ESPRESSO.
• MD simulations include a GUI for LAMMPS and can analyze physical movements of atoms at finite temperature, thermal, optical and mechanical properties.
• Pricing ranges from pay-per-use of $0.25 per core hour to UNLIMITED plans.

References:

1. Materials Square Brochure

Answer 2

Amazon Web Services (AWS) and Google Cloud can be used for the running simulations and I've used both. I applied to a few grant programs a while back where part of the award included AWS or Google Cloud credits. I think the credits were donated to the funding agencies by the respective companies to encourage more use of those services in science and engineering.

Anyway, it was pretty easy to put together an environment to run LAMMPS on both these services. The cores themselves were not that fast, but the environments were well suited for running replica exchange calculations with large numbers of replicas (50 to hundreds). The communication between the "virtual CPUs (vCPUs)" was sufficiently fast that I didn't see any performance degradation between serial runs and replica exchange. That said, exchanges were only attempted every 1000 steps (~3 minutes of real time) in my simulations, so communication would have had to be really slow to see a major impact.

I haven't used them since preparing the those grant applications since the cost seemed somewhat high compared to buying a workstation computer and running it for years.

Answer 3

Molecular Quantum Chemistry via WebMO on Google Cloud

It's fairly easy to run Linux-based programs on Google Cloud. The WebMO interface gives a tutorial on installing on GCP including molecular quantum chemistry programs.

Supported tools include open source and "free binary" programs, for example:

• Psi4
• Orca
• GAMESS
• NWChem
• Mopac

As mentioned in another answer, costs are higher than a dedicated workstation or cluster, but if you're performing occasional calculations, it might be worthwhile.

Answer 4

Hetzner

If you know you will continuously need a powerful computer for the next 1-2 years, consider renting a bare-metal server from Hetzner. You can rent a 32 core 2nd generation AMD EPYC server with 256 GB of RAM and a fast SSD for about 200 EUR/month. No traffic fees.

You get remote access to a blank machine, and you are free to install whatever OS and software you need, no virtualization overhead or other restrictions.

There are other configurations to choose from, but the AMD boxes are usually the better value.

PS: The 32 core box I mentioned is seriously powerful when it comes to CPU compute. You can expect Linpack perfomance in excess of 1 TFLOPS, as the 24 core model of the same family can do 1 TFLOPS +/- 10%.

One thing you need to watch out for is the number of memory modules! If you rent a server with less than half of the maximum memory capacity, some of the memory channels will be left unpopulated, which may have a negative impact on performance.

So for the EPYC server I would recommend exactly 256 GB of RAM, or if (for your applications) that is not enough to go with 32 cores, then 512 GB would be the next good option.

Answer 5

Amazon Elastic Compute Cloud (Amazon EC2)

AWS Free Tier (one year free trial) includes 750 hours per month for 12 months of t2.micro (1 vCPU, 1GB Memory) or t3.micro (2 vCPU, 1GB Memory) instance depending on region.

You can choose from various instance types that is best suitable for your use case.

There are a few pricing models:

1. On-Demand

   • You pay for only what you use (by the hour or the second depending on which instances you run).

   • E.g. Hourly on-demand rates for general purpose instances with 8 vCPU are ranged from 0.2 to 0.7 USD per hour.

     

2. Spot instances

   • Spot instance prices are adjusted gradually based on supply and demand trends, one can use spot instances and save up to 90% off the on-demand price.

3. Reserved Instances

   • Commit to using EC2 for a 1 or 3 year term and save up to 75% comparing to on-demand price

4. Dedicated Hosts