perf(graph,mpi): ⚡ derive the exchange layouts and store only occupied world slots - #270

Merged

robertodr merged 48 commits into

mainfrom

perf/sparse-slots-and-funnel

Aug 25, 2026

Commits on Aug 24, 2026

perf(evolution): :zap: stop storing the cross-rank D range twice
diagonal-hamiltonian
committed
cadbea1
perf(evolution): :zap: derive the exchange layouts instead of retaining them
diagonal-hamiltonian
committed
c33cc2f
perf(evolution): :zap: derive the recv layout too, instead of caching the transpose

diagonal-hamiltonian
and
claude
committed
16227b0
feat(graph): :chart_with_upwards_trend: separate world-slot metadata from traffic in the graph breakdown
diagonal-hamiltonian
committed
ab94124
perf(graph): :zap: store only the world slots that carry traffic
diagonal-hamiltonian
committed
a613887
perf(graph): :zap: scatter the derived exchange layout over the occupied slots
diagonal-hamiltonian
committed
94914f6
fix(mpi): :bug: check the exchange layout width outside the symmetry gate
diagonal-hamiltonian
committed
0975bf2
refactor(mpi): :recycle: gate the exchange symmetry audit at build time
diagonal-hamiltonian
committed
3d26252
test(graph): :white_check_mark: move the exchange-layout oracle out of the library
diagonal-hamiltonian
committed
91d330f
refactor(graph): :broom: drop the two permanently-zero breakdown diagnostics
diagonal-hamiltonian
committed
580b1b4
refactor(graph): :broom: delete layer_exchange_layout_storage_bytes
diagonal-hamiltonian
committed
6d10d6c
fix(test): :bug: assert the occupied-slot layout rule, not a sum of field widths
diagonal-hamiltonian
committed
8df13ec
fix(graph): :bug: refuse an in-block that does not fit inside its endpoint list
diagonal-hamiltonian
committed
73d6fbe
docs: :memo: record the distributed-graph memory rule and the build options
diagonal-hamiltonian
committed
135026b
style: :art: apply clang-format to the lines this branch added
diagonal-hamiltonian
committed
5ea1085
refactor(evolution): :broom: drop the rank arguments the self-slot lookup made unused
diagonal-hamiltonian
committed
44ed589
fix(graph): :bug: assert the occupied-slot size rule on both term-index widths
diagonal-hamiltonian
committed
29f7fb6
test(graph): :white_check_mark: cover graph_memory_breakdown() from Python
diagonal-hamiltonian
committed
b9ab380
docs: :memo: record the both-widths and ledger-contract conventions
diagonal-hamiltonian
committed
517c745
perf(mpi): :zap: pay the hybrid staging matrix on its owners, not on partition 0
diagonal-hamiltonian
committed
d91adf3
test(mpi): :white_check_mark: pin the hybrid staging tiling with pairwise counts
diagonal-hamiltonian
committed
be2d876
docs(mpi): :memo: correct what Phase P0 actually moved
diagonal-hamiltonian
committed
037b003
test(mpi): :memo: name the mutants the pairwise counts actually catch
diagonal-hamiltonian
committed
0b049dc
fix(graph): :bug: check the occupied-slot counts instead of casting them
diagonal-hamiltonian
committed
60291df
refactor(graph): :broom: delete the rank-count check that cannot fail
diagonal-hamiltonian
committed
73553a8
docs(graph): :memo: the layer cores hold no eval-time cache
diagonal-hamiltonian
committed
7eb2910
perf(graph): :zap: format the exchange overflow label only when it throws
diagonal-hamiltonian
committed
84ffbc8
refactor(graph): :broom: drop the scale-1 build-time derivation
diagonal-hamiltonian
committed
976f602
perf(evolution): :zap: index the derivative snapshot by occupied position
diagonal-hamiltonian
committed
755435a
refactor(mpi): :broom: size the hybrid send messages from the column sums
diagonal-hamiltonian
committed
eaf04ba
refactor(mpi): :broom: pack from the count row pack_off_ was built from
diagonal-hamiltonian
committed
e63d284
refactor(mpi): :broom: reuse col_sum_ as pass B's prefix and drop run_
diagonal-hamiltonian
committed
eaace17
refactor(graph): :fire: drop the LayerTraversal members nothing calls

diagonal-hamiltonian
and
claude
committed
0869671
docs: :memo: cut the branch's comments to what a reviewer can check

diagonal-hamiltonian
and
claude
committed
7d61602
test(mpi): :white_check_mark: prove the symmetry audit fires, and refuse a rank list that cannot

diagonal-hamiltonian
and
claude
committed
234ad2d
test(graph): :bug: declare the overflow boundary count instead of materialising it

diagonal-hamiltonian
and
claude
committed
b253178
refactor(mpi): 🔥 drop the exchange symmetry audit and its build option

diagonal-hamiltonian
and
claude
committed
f7c74de
refactor(graph): 🚨 throw a dedicated type for the slot layout invariants

diagonal-hamiltonian
and
claude
committed
f682ad7
refactor: 🎨 answer the static analysers on this branch's new code

diagonal-hamiltonian
and
claude
committed
f718b11
docs: ✂️ cut this branch's comments back

diagonal-hamiltonian
and
claude
committed
dc1f781
docs: ⏪ drop this branch's documentation additions

diagonal-hamiltonian
and
claude
committed
955c6ff
test(cmake): 🔥 drop the multi-rank rank-list guard

diagonal-hamiltonian
and
claude
committed
1d9fef3
style(sonar): 🎨 drop the blank line left by removing the layercore block

diagonal-hamiltonian
and
claude
committed
4491fc8
Merge branch 'main' into perf/sparse-slots-and-funnel
diagonal-hamiltonian
authored
a411427
Merge branch 'main' into perf/sparse-slots-and-funnel
diagonal-hamiltonian
authored
5edeb53

Commits on Aug 25, 2026

Merge branch 'main' into perf/sparse-slots-and-funnel
robertodr
authored
2970627
Remove graph memory section from parallelism docs
diagonal-hamiltonian
authored
92814a6
refactor(evolution,graph): 🎨 apply review comments

diagonal-hamiltonian
and
claude
committed
3a87c66