Creating multigrid hierarchy (e.g., `ruge_stuben`) allocates too much memory

[ma-sadeghi]

For a test problem (Laplace equation on a 200^3 grid), ruge_stuben allocates ~8 GB of memory. For comparison, LinearSolve.KrylovJL_CG allocates ~90 MB.

Is this normal / can it be optimized?

Thanks!

[termi-official]

Sorry, took me a while to come back to this. The majority of the allocations can here be attributed to the matrix triple products (called RAP) necessary to generate the coarser representations. My best attempt so far was to reduce the number of allocations by 50%, but at the cost of runtime performance (almost doubled for CSC matrices). It is also possible to improve the runtime here by about 25% by switching to CSR matrices with a faster, custom RAP implementation. I have sketched some here https://gist.github.com/termi-official/d131c437fdfcde23f2f0f3cce60fae20 , but they need more curation and we need to think more carefully how to integrate these properly.

In the new release (v.1.2) you can now track down where the time is spend by TimerOutputs.enable_debug_timings(AlgebraicMultigrid). This should help tracking down performance issues in future issues.