Status: Planned · Connect distributed sparse operations to communication costs and reproducible scaling measurements.

Question

How much of an iterative solver’s time is spent computing, communicating, and waiting for global reductions?

Planned sequence

1. Specify matrix ownership, partitioning, and a serial reference result.
2. Implement halo exchange for distributed sparse matrix-vector products.
3. Add dot products and global reductions needed by a Krylov iteration.
4. Compare blocking and nonblocking communication after verifying correctness.
5. Measure strong and weak scaling, then connect the results to domain decomposition.

Evidence to record

Record hardware, ranks, threads, placement and binding, compiler flags, MPI version, matrix size and nonzeros, repetitions, accuracy checks, and exact build/run commands. Separate compute, halo exchange, and reduction timings.

Current evidence

This is a planned project. MPI/CMake source, timing data, and scaling figures will be linked when available.

Study prerequisites

Sparse linear systems → Parallel / PDE connection

Linked notes

No project notes have been published yet. The steps above describe planned work.

Linked paper notes

No paper notes linked yet.

All projects · Research directions