[djlampert]

Hello--I am interested in using SU2 to develop a 3-D lake simulation of fluid flow and heat transfer to run on a cluster. However, while I have some experience in numerical modeling, and a lot of practice writing Python code, I haven't used SU2 or extensively used a cluster previously. I have to write a short proposal for an educational allocation on the cluster, and they want to know the number of node-hours and the balance between x86 CPUs and Kepler GPUs. I really don't know exactly how to make a good estimate of the node-hours, or which processor would work better, but I need to submit this quickly. Any advice would be welcome.

[hlk]

SU2 is generally compiled for CPUs, making it efficient (or work at all) on GPUs (while a potentially interesting project) would be a hefty task on its own.
In terms of the number of hours, if you don't have access to test on the cluster, then what you can do is to run a small case on your own computer to get a basic idea of how much time is needed, scaling up by how large the mesh you'll actually use will be and how many simulations you expect. Often SU2 is estimated to stop scaling well after 5000 nodes per processor, but this depends on the cluster architecture.
Clusters sometimes give a small-scale allocation to allow people to test for this reason.

Developing your simulations at a small scale/coarse mesh that can run on your local computer, although not expected to generate accurate results, is a very good idea both to get an idea of how much time you will need on the cluster, and to work out any problems without using up the more finite cluster time.

Note that once you get onto the cluster, if the existing automatically-used parallel commands (mpiexec or mpirun) aren't compatible with their system you'll need to set the SU2_MPI_COMMAND option.

Originally Posted by djlampert

Hello--I am interested in using SU2 to develop a 3-D lake simulation of fluid flow and heat transfer to run on a cluster. However, while I have some experience in numerical modeling, and a lot of practice writing Python code, I haven't used SU2 or extensively used a cluster previously. I have to write a short proposal for an educational allocation on the cluster, and they want to know the number of node-hours and the balance between x86 CPUs and Kepler GPUs. I really don't know exactly how to make a good estimate of the node-hours, or which processor would work better, but I need to submit this quickly. Any advice would be welcome.

[djlampert]

Hi Heather. Thanks so much for the advice! So it sounds like we can't use GPUs, so I will request 100% of the allocation be applied to CPUs. I have a colleague that raves about the performance of GPUs for his work, but I don't really understand how they are different in a fundamental way.

We will definitely be trying to use coarse-mesh, simple implementations of the model for the first part of the project on a desktop and then on the cluster. I have some funding to get a good desktop workstation that would be helpful for this also--any thoughts on a good value for SU2-based simulations for $3-5,000?