Parallel scalability benchmarks? about veros HOT 2 CLOSED

Nice work with Veros!

Thanks 😁

Can Veros run in distributed memory mode (beyond a single node)?

Currently, no. It is our philosophy to handle all parallelism in the backend, and until we implement bindings to a suitable distributed library, we will only support shared memory architectures. I am currently evaluating whether we can modify dask.array to fit our needs, but it will take a while.

Have you done any parallel scalability benchmarks -- even if only for shared memory?

Yes! Those were the ones you looked at. I can understand that they may seem confusing if you're not familiar with our backends. NumPy is a sequential library, so the number of cores doesn't matter, and architecture 1 has better single-core performance, so NumPy runs a bit faster (the same applies for the sequential Fortran backend). Bohrium and Fortran + MPI are parallelized, that's why they run significantly faster on the 24-core architecture.

The takeaway message from those benchmarks is that NumPy performs better than Bohrium for small setups, but for large problems the benefits are huge (and performance is in the same order of magnitude as the MPI-parallelized Fortran library).

If you have a suggestion how the benchmarks could be explained better, I'd encourage you to submit a PR :)

from veros.

OK, now I see what you're saying. Basically, on diagram II, the horizontal lines of Bh GPU and Bh CPU in the 1e3 - 1e6 elements range implies scalability...

I think it would be helpful to see t(1) / t(n) for n = 1, 2, 4, ..., 24 cores.

from veros.