perf(mpi): token protocol for the neighbour mesh exchanges

[gouarin]

• I have installed pre-commit locally and use it to validate my commits.
• The PR title follows the conventional commits convention.
  Available tags: 'build', 'chore', 'ci', 'docs', 'feat', 'fix', 'perf', 'refactor', 'revert', 'style', 'test'
• This new PR is documented.
• This new PR is tested.

Description

Rebased on main after #490/#491.

The graduation half of this PR is dropped on rebase. #482 replaced the graduation fixed point by the single-pass algorithm, which exchanges one level array per level (exchange_level_mpi) and has no repeated identical send to elide - update_subdomains_mpi and the fixed-point loop the token protocol hooked into no longer exist. What remains here is the mesh-exchange half.

Every mesh construction serialized and sent the full mesh state to each neighbour three times (subdomain, cells, then the whole mesh with every derived mesh id) - even when the receivers already held identical data from the previous construction.

This PR replaces those exchanges by a token protocol: a one-int header announces whether the serialized payload follows; when it does not, the receiver keeps its cached copy of the neighbour's data. A sender only skips the payload when the receivers provably hold identical data:

• mesh exchanges (update_neighbour_subdomain, update_meshid_neighbour, update_mesh_neighbour): cells geometrically identical to the reference mesh (same_cells, from perf(mesh): detect the adaptation fixed point before constructing the new mesh #476), same neighbour set and - for the whole-mesh exchange, whose derived ids depend on the neighbours' cells - a token received from every neighbour in the cells exchange. find_neighbourhood now carries the previously exchanged neighbour meshes over to the rebuilt neighbourhood; they are the receive-side cache.

The exchanges stay collective over the (symmetric) neighbourhood, so the communication pattern remains matched.

Measurements on advection_2d (Tf 0.1, Apple M, MPICH + libfabric, interleaved A/B vs #476), taken before the rebase, i.e. on the pre-#482 graduation and with both halves of the original change in place. They are kept for reference; the mesh-exchange half alone on today's main has not been re-measured:

Config	before	after
8 ranks, libfabric sockets provider	167 s	120 s (-28%)
8 ranks, FI_PROVIDER=tcp	7.6-7.7 s	7.4-7.5 s (-3%)
4 ranks	3.09-3.14 s	3.03-3.10 s (-2%)
serial	parity	parity

Tokens cut the payload volume, which pays hugely on expensive transports and grows with rank count and mesh size. On a sane transport at low rank counts, the remaining MPI cost of the adaptation loop is the number of synchronous exchange rounds - addressed in #478.

Related issue

None.

How has this been tested?

Outputs are unchanged by construction: a token replaces the payload only when the receiver already holds an identical copy.

Before the rebase (on the pre-#482 base):

• h5diff clean vs perf(mesh): detect the adaptation fixed point before constructing the new mesh #476 in serial and at 4 and 8 ranks (demos/FiniteVolume/advection_2d.cpp, Tf 0.5);
• tests/: 350/350 serial; tests/mpi/: 24/24 at 2/3/4 ranks, including the load-balancing migration tests which exercise the carried-over neighbour caches;
• all demos build (serial and MPI).

After the rebase, the rebased branch is verified by CI.

Code of Conduct

By submitting this PR, you agree to follow our Code of Conduct

• I agree to follow this project's Code of Conduct

[codacy-production]

Up to standards ✅

🟢 Issues 0 issues

Results:
0 new issues

View in Codacy

🟢 Metrics -8 duplication

Metric	Results
Duplication	-8

View in Codacy

_{NEW Get contextual insights on your PRs based on Codacy's metrics, along with PR and Jira context, without leaving GitHub. Enable AI reviewer
TIP This summary will be updated as you push new changes.}