slurm VS mfem

I'm trying to have Slurm automatically switch partitions to a specific one via the job_sutmit.lua plugin whenever our users request strictly more than 8 cpus. But trying to extract or calculate ahead of time how many cpus will be allocated or requested isn't trivial (to me). Are there attributes in job_submit that could help out with this task? For example, I don't see any job->desc.ntasks attribute in https://github.com/SchedMD/slurm/blob/master/src/plugins/job_submit/lua/job_submit_lua.c. Any information or documentation on how to leverage job_submit.lua would be appreciated.

Do you have a reason to use kubernetes besides it’s the $CURRENT tech? Why not stick with what you’re already familiar with (batch job managers) and use SLURM, a workload and resource manager, like many others in HPC? Do the researchers need to schedule against Nvidia GPU resources now or in the future? Nvidia themselves recommend SLURM.

Quantum computing and supercomputers are two different things. Quantum computers are currently an area of research, there isn't a version ready for use apart from some prototypes, and it will probably stay that way for while. Also, quantum computing will most likely not be a completly new architecture, that all of the chips we use will adopt, but an addition to current chipsets for some important but special tasks. Supercomputers, or HPC (High performance clusters) are classic computers, just that they are huge. They use derivatives of "off-the-shelf", but high end, hardware. There is a lot of interesting work in designing such systems, a lot of challenging problems in distributed systems theory, but they aren't a complete detached industry. Using them for work, not designing them, doesn't require a EECS type degree, they guy who sit's next to me in the office, uses a supercomputer to predict protein folding, he is by training a doctor and now does computational microbiology. The applications for massive compute power (often times "just brute force the solution instead of spending years in the lab") are almost endless, but to use them it's not that important to understand the full details of how they are constructed, domain knowledge in the application domain is much more important. If you know how your cluster is structured, and knowledge of slurm etc. will enable you to use the supercomputer just fine, again, they aren't that different from regular computers, just that you workstation might have 1 CPU and your supercomputer has 500. Hiding this complexity is done by slurm or any other resource manager. It's open source as well :) https://github.com/SchedMD/slurm

SLURM (distributed by OpenHPC) If you have shared storage then this is the industry standard solution that is both open source and free (extremely popular in the top 500 list). You can pair this with a high speed network or not depending on your research workloads.

MFEM (mfem.org) is one of the most active open-source partial differential equation (PDE) solver projects in recent years. Although this open source package is positioned as a lightweight and scalable C++ library, it provides higher-order finite element spaces, supports mixed elements, discontinuous Galerkin elements, isogeometric analysis methods, and more. In particular, it has great advantages in high-performance computing, not only supports message passing interface (MPI) parallelism and shared memory parallelism (OpenMP), but also has good strength in GPU parallel computing. The built-in post-processing program GLVis can easily read and display the result files. The BSD-3 license is also extremely friendly to developers. Recently, MFEM supports Python programming, which makes the library more convenient for various types of researchers.

If you are looking for something more numerically intensive, check out MFEM: https://mfem.org. Lots of examples in the gallery there; lots of cool higher order curved meshes and physics problems to play with. Probably a higher learning curve than some of your other options.

If you're interested in a FEM framework designed by computer scientists and mathematicians, MFEM has treated me well. Maybe just something to compare to.

We currently have several solvers of our own, as well as interfacing with MPI capable solvers via mfem. We are currently a pre and post processor for several NASTRAN based solvers using our meshing free method.

MFEM

I’m working on MFEM. It is a Finite Element Method (FEM) library. Some of the work would require knowledge of the FEM but there’s a lot of stuff like implementing operators in vector classes that you could do too!